On Tue, Oct 17, 2017 at 09:47:05AM -0700, Rodrigo Vivi wrote:
> On Tue, Oct 17, 2017 at 03:44:21PM +0000, Ville Syrjälä wrote:
> > On Tue, Oct 03, 2017 at 12:06:08AM -0700, Rodrigo Vivi wrote:
> > > From: "Kahola, Mika" <mika.kahola@xxxxxxxxx>
> > >
> > > Display Voltage and Frequency Switching (DVFS) is used to adjust the
> > > display voltage to match the display clock frequencies. If voltage is
> > > set too low, it will break functionality. If voltage is set too high,
> > > it will waste power. Voltage level is selected based on CD clock and
> > > DDI clock.
> > >
> > > The sequence before frequency change is the following and it requests
> > > the power controller to raise voltage to maximum
> > >
> > > - Ensure any previous GT Driver Mailbox transaction is complete.
> > > - Write GT Driver Mailbox Data Low = 0x3.
> > > - Write GT Driver Mailbox Data High = 0x0.
> > > - Write GT Driver Mailbox Interface = 0x80000007.
> > > - Poll GT Driver Mailbox Interface for Run/Busy indication cleared (bit 31 = 0).
> > > - Read GT Driver Mailbox Data Low, if bit 0 is 0x1, continue, else restart the sequence.
> > > Timeout after 3ms
> > >
> > > The sequence after frequency change is the following and it requests
> > > the port controller to raise voltage to the requested level.
> > >
> > > - Write GT Driver Mailbox Data Low
> > > * For level 0, write 0x0
> > > * For level 1, write 0x1
> > > * For level 2, write 0x2
> > > * For level 3, write 0x3
> > > - Write GT Driver Mailbox Data High = 0x0.
> > > - Write GT Driver Mailbox Interface = 0x80000007.
> > >
> > > For Cannonlake, the level 3 is not used and it aliases to level 2.
> > >
> > > v2: reuse Paulo's work on cdclk. This patch depends on Paulo's patch
> > > [PATCH 02/12] drm/i915/cnl: extract cnl_dvfs_{pre,post}_change
> > > v3: (By Rodrigo): Remove duplicated commend and fix typo on Paulo's name.
> > > v4: (By Rodrigo): Rebase on top "Unify and export gen9+ port_clock calculation"
> > > The port clock calculation here was only addressing DP, so let's reuse
> > > the current port calculation that is already in place without any duplication.
> > > Alos fix portclk <= 594000 instead of portclk < 594000.
> > >
> > > Cc: Paulo Zanoni <paulo.r.zanoni@xxxxxxxxx>
> > > Cc: Ville Syrjälä <ville.syrjala@xxxxxxxxxxxxxxx>
> > > Signed-off-by: Kahola, Mika <mika.kahola@xxxxxxxxx>
> > > Signed-off-by: Rodrigo Vivi <rodrigo.vivi@xxxxxxxxx>
> > > ---
> > > drivers/gpu/drm/i915/intel_dpll_mgr.c | 30 ++++++++++++++++++++++--------
> > > 1 file changed, 22 insertions(+), 8 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/i915/intel_dpll_mgr.c b/drivers/gpu/drm/i915/intel_dpll_mgr.c
> > > index a2a3d93d67bd..6030fbafa580 100644
> > > --- a/drivers/gpu/drm/i915/intel_dpll_mgr.c
> > > +++ b/drivers/gpu/drm/i915/intel_dpll_mgr.c
> > > @@ -1966,10 +1966,23 @@ static const struct intel_dpll_mgr bxt_pll_mgr = {
> > > .dump_hw_state = bxt_dump_hw_state,
> > > };
> > >
> > > +static int cnl_get_dvfs_level(int cdclk, int portclk)
> > > +{
> > > + if (cdclk == 168000 && portclk <= 594000)
> > > + return 0;
> > > + else if (cdclk == 336000 && portclk <= 594000)
> > > + return 1;
> > > + else
> > > + return 2;
> > > +}
> > > +
> > > static void cnl_ddi_pll_enable(struct drm_i915_private *dev_priv,
> > > struct intel_shared_dpll *pll)
> > > {
> > > uint32_t val;
> > > + int ret;
> > > + int level;
> > > + int cdclk, portclk;
> > >
> > > /* 1. Enable DPLL power in DPLL_ENABLE. */
> > > val = I915_READ(CNL_DPLL_ENABLE(pll->id));
> > > @@ -2006,11 +2019,9 @@ static void cnl_ddi_pll_enable(struct drm_i915_private *dev_priv,
> > > /*
> > > * 5. If the frequency will result in a change to the voltage
> > > * requirement, follow the Display Voltage Frequency Switching
> > > - * Sequence Before Frequency Change
> > > - *
> > > - * FIXME: (DVFS) is used to adjust the display voltage to match the
> > > - * display clock frequencies
> > > + * (DVFS) Sequence Before Frequency Change
> > > */
> > > + ret = cnl_dvfs_pre_change(dev_priv);
> > >
> > > /* 6. Enable DPLL in DPLL_ENABLE. */
> > > val = I915_READ(CNL_DPLL_ENABLE(pll->id));
> > > @@ -2028,11 +2039,14 @@ static void cnl_ddi_pll_enable(struct drm_i915_private *dev_priv,
> > > /*
> > > * 8. If the frequency will result in a change to the voltage
> > > * requirement, follow the Display Voltage Frequency Switching
> > > - * Sequence After Frequency Change
> > > - *
> > > - * FIXME: (DVFS) is used to adjust the display voltage to match the
> > > - * display clock frequencies
> > > + * (DVFS) Sequence After Frequency Change
> > > */
> > > + if (ret == 0) {
> > > + cdclk = dev_priv->cdclk.hw.cdclk;
> > > + portclk = intel_ddi_port_clock(dev_priv, pll->id);
> > > + level = cnl_get_dvfs_level(cdclk, portclk);
> > > + cnl_dvfs_post_change(dev_priv, level);
> >
> > This isn't how I imagined we'd handle this. What I was after is
> > pre-computing the correct "dfvfs level" (or what I'd probably just call
> > "voltage" or something like that), and then we'd just let the normal
> > cdclk programming sequence do its thing.
> >
> > Is it even safe to do this with other pipes/ports enabled? I would
> > have assumed no, but maybe I'm mistaken?
>
> Well, this is how the spec is written. So I assume it is safe to change
> that at that point.
>
> Also we enable CDCLK during boot while we have no way to compute the
> link rate. So I don't see a better way of hadling that.
We do a full state readout, which should tell us what voltage we need.
And if the cdclk isn't even enabled, then obviously we can't have any
pipes running either, so we can safely enable it with any frequency
even before the full state readout.
> So, Art is it ok in the way that I'm proposing here:
>
> on core display init:
>
> - dvfs-pre;
> - enable cdclk;
> - dvfs-post considering cdclk only and link rate 0;
>
> on pll init:
>
> - consider cdclk and link rate and if desired level is different from what is currently set: dvfs-pre
> - enable pll
> - consider cdclk and link rate and if desired level is different from what is currently set: dvfs-post with new level.
>
> This is how I read the spec, but please let me know what I'm missing.
Hmm. OK, so maybe we can change the voltage while things are up and
running.
That said, the voltage level is a global thing, so whenever we change it
we must consider the state of all ports, and the cdclk. This looks like
it only considers the current port. Thus if we first enable a port that
needs high voltage, and later a port that can make due with a lower
voltage we'll end up with something not working.
We migth also end up changing the voltage multiple times if we enable
multiple ports with one atomic operation. If things went via the normal
cdclk path we'd just do the voltage change the once.
Additonally I don't like that the code is getting the current clock by
means of readout in the middle the modeset. With atomic the design is
pretty much to pass all state in explicitly, and readout is only used
to take over the from the BIOS, and to verify that we programmed the
hardware correctly.
>
> Maybe I'm missing the link rate for eDP during core display init like we are currently kind of missing for SKL?
Hmm. What exactly are we missing?
--
Ville Syrjälä
Intel OTC
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
