[Intel-gfx] [PATCH] drm/i915: Reset dpll_hw_state when selecting a new pll on hsw

[Daniel Vetter]

On Wed, Oct 14, 2015 at 04:58:55PM +0300, Ander Conselvan De Oliveira wrote:
> On Wed, 2015-10-14 at 14:44 +0200, Daniel Vetter wrote:
> > On Wed, Oct 14, 2015 at 11:21:46AM +0300, Ander Conselvan De Oliveira wrote:
> > > On Tue, 2015-10-13 at 16:08 +0200, Daniel Vetter wrote:
> > > > On Tue, Oct 13, 2015 at 04:00:37PM +0200, Maarten Lankhorst wrote:
> > > > > Op 13-10-15 om 15:58 schreef Daniel Vetter:
> > > > > > On Tue, Oct 13, 2015 at 03:43:28PM +0200, Maarten Lankhorst wrote:
> > > > > > > Op 13-10-15 om 15:35 schreef Daniel Vetter:
> > > > > > > > On Tue, Oct 13, 2015 at 03:18:16PM +0200, Maarten Lankhorst wrote:
> > > > > > > > > Op 23-09-15 om 17:34 schreef Gabriel Feceoru:
> > > > > > > > > > Using 2 connectors (DVI and VGA) will cause wrpll to be set for
> > > > > > > > > > INTEL_OUTPUT_HDMI but never reset if switching to INTEL_OUTPUT_VGA
> > > > > > > > > > 
> > > > > > > > > > Supresses errors like these:
> > > > > > > > > > [drm:intel_pipe_config_compare [i915]] *ERROR* mismatch in dpll_hw_state.wrpll
> > > > > > > > > > 
> > > > > > > > > Looks like a good idea to always zero it.
> > > > > > > > Except that we still have a bunch of cases where we recompute clock state
> > > > > > > > but only partially. Can we just move them all up into a common place
> > > > > > > > please? That would also catch cases where we simply forget to fill this
> > > > > > > > out at all.
> > > > > > > > 
> > > > > > > > One case I noticed is edp in skl_ddi_pll_select, but there's probably
> > > > > > > > more.
> > > > > > > > 
> > > > > > > Something like below, with all the memsets for dpll_hw_state removed?
> > > > > > I think this will blow up since we recompute clock state only when
> > > > > > needs_modeset is true. So needs a bit more intelligence in deciding when
> > > > > > to clear it I think.
> > > > > Oops you're right. Maybe intel_modeset_clear_plls because that's where all the clock state
> > > > > belongs?
> > > > 
> > > > Yeah that might be an even better place, in the loop after the continue;
> > > > statement.
> > > 
> > > The reason I didn't put the memset there in the first place was the way we calculate plls for DP
> > > with DDI platforms. In that case, ddi_pll_sel is setup from the encoder_config instead of
> > > compute_clock, so a memset ends up clearing the new pll config.
> > 
> > Hm, I forgot about this split totally. And there seems to be a giant mess
> > going on here:
> > 
> > In our top-level intel_atomic_check we have 4 parts to compute state:
> > 1. drm_atomic_helper_check_modeset
> > 2. intel_modeset_pipe_config
> > 3. intel_modeset_checks
> > 4. drm_atomic_helper_check_planes
> > 
> > We recalculate clocks (by calling dev_priv->display.crtc_compute_clock)
> > in 1., way ahead of anything else in intel_crtc_atomic_check. That looks
> > very suspcious since it means only very later on (in the loop that does
> > 2.) do we even decide whether we need to do a full modeset or not.
> > 
> > So what I had in mind is that we clear clocks in
> > intel_modeset_pipe_config, before we call any of the callbacks. That makes
> > sure that when we decided to do a modeset, we do recompute the clocks
> > correctly.
> 
> I had a suspicion this would interact badly with how we "cancel" the modeset if the pipe config
> didn't changed, just after the call to intel_modeset_pipe_config(). It turns out there's an issue
> there already.
> 
> There are two possibilities for the dpll_hw_state value after the new pipe_config is calculated. It
> may have the new values already for DP in HSW/BDW and eDP in SKL or it may still have the old value.
> In the latter case the new value is only calculated in .crtc_clock(), after we already compared the
> old and new configs and may have decided to skip the modeset.
> 
> But doing the memset() in intel_modeset_pipe_config() would be find as long as we don't change our
> minds about doing a modeset later.

It's more annoying since my analysis is all wrong: intel_crtc_atomic_check
is called from drm_atomic_helper_check_planes, i.e. step 4 not step 1.
It'll all work out I think if we memset it in intel_modeset_pipe_config.
The caveat is that we need to move the clock recomputation into
intel_modeset_pipe_config too (which is better, since then we'll have more
accurate state to decided whether we'll fastboot or not).

And then intel_modeset_clear_plls would really just update the global pll
setup (and would be really good to rename it to
intel_modeset_compute_shared_dpll or whatever).

Thoughts?
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch