[Intel-gfx] [PATCH v26 3/9] drm/i915: Track active_pipes in bw_state

Lisovskiy, Stanislav stanislav.lisovskiy at intel.com
Thu Apr 30 11:48:35 UTC 2020


On Thu, Apr 30, 2020 at 02:40:37PM +0300, Ville Syrjälä wrote:
> On Thu, Apr 30, 2020 at 02:29:51PM +0300, Lisovskiy, Stanislav wrote:
> > On Thu, Apr 30, 2020 at 02:22:02PM +0300, Ville Syrjälä wrote:
> > > On Thu, Apr 30, 2020 at 02:07:02PM +0300, Lisovskiy, Stanislav wrote:
> > > > On Thu, Apr 30, 2020 at 01:55:59PM +0300, Ville Syrjälä wrote:
> > > > > On Thu, Apr 30, 2020 at 01:47:02PM +0300, Lisovskiy, Stanislav wrote:
> > > > > > On Thu, Apr 30, 2020 at 01:32:17PM +0300, Ville Syrjälä wrote:
> > > > > > > On Thu, Apr 30, 2020 at 01:05:15PM +0300, Lisovskiy, Stanislav wrote:
> > > > > > > > On Thu, Apr 30, 2020 at 12:21:04PM +0300, Ville Syrjälä wrote:
> > > > > > > > > On Thu, Apr 23, 2020 at 10:58:56AM +0300, Stanislav Lisovskiy wrote:
> > > > > > > > > > We need to calculate SAGV mask also in a non-modeset
> > > > > > > > > > commit, however currently active_pipes are only calculated
> > > > > > > > > > for modesets in global atomic state, thus now we will be
> > > > > > > > > > tracking those also in bw_state in order to be able to
> > > > > > > > > > properly access global data.
> > > > > > > > > > 
> > > > > > > > > > Signed-off-by: Stanislav Lisovskiy <stanislav.lisovskiy at intel.com>
> > > > > > > > > > ---
> > > > > > > > > >  drivers/gpu/drm/i915/display/intel_bw.h |  3 +++
> > > > > > > > > >  drivers/gpu/drm/i915/intel_pm.c         | 15 ++++++++++-----
> > > > > > > > > >  2 files changed, 13 insertions(+), 5 deletions(-)
> > > > > > > > > > 
> > > > > > > > > > diff --git a/drivers/gpu/drm/i915/display/intel_bw.h b/drivers/gpu/drm/i915/display/intel_bw.h
> > > > > > > > > > index d6df91058223..898b4a85ccab 100644
> > > > > > > > > > --- a/drivers/gpu/drm/i915/display/intel_bw.h
> > > > > > > > > > +++ b/drivers/gpu/drm/i915/display/intel_bw.h
> > > > > > > > > > @@ -26,6 +26,9 @@ struct intel_bw_state {
> > > > > > > > > >  
> > > > > > > > > >  	unsigned int data_rate[I915_MAX_PIPES];
> > > > > > > > > >  	u8 num_active_planes[I915_MAX_PIPES];
> > > > > > > > > > +
> > > > > > > > > > +	/* bitmask of active pipes */
> > > > > > > > > > +	u8 active_pipes;
> > > > > > > > > >  };
> > > > > > > > > >  
> > > > > > > > > >  #define to_intel_bw_state(x) container_of((x), struct intel_bw_state, base)
> > > > > > > > > > diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> > > > > > > > > > index 7e15cf3368ad..f7249bca3f6f 100644
> > > > > > > > > > --- a/drivers/gpu/drm/i915/intel_pm.c
> > > > > > > > > > +++ b/drivers/gpu/drm/i915/intel_pm.c
> > > > > > > > > > @@ -3874,6 +3874,7 @@ static int intel_compute_sagv_mask(struct intel_atomic_state *state)
> > > > > > > > > >  	struct intel_bw_state *new_bw_state = NULL;
> > > > > > > > > >  	const struct intel_bw_state *old_bw_state = NULL;
> > > > > > > > > >  	int i;
> > > > > > > > > > +	bool active_pipes_calculated = false;
> > > > > > > > > >  
> > > > > > > > > >  	for_each_new_intel_crtc_in_state(state, crtc,
> > > > > > > > > >  					 new_crtc_state, i) {
> > > > > > > > > > @@ -3883,6 +3884,12 @@ static int intel_compute_sagv_mask(struct intel_atomic_state *state)
> > > > > > > > > >  
> > > > > > > > > >  		old_bw_state = intel_atomic_get_old_bw_state(state);
> > > > > > > > > >  
> > > > > > > > > > +		if (!active_pipes_calculated) {
> > > > > > > > > > +			state->active_pipes = new_bw_state->active_pipes =
> > > > > > > > > 
> > > > > > > > > I don't think we should touch state->active_pipes here.
> > > > > > > > 
> > > > > > > > Well, that was my question actually here as well. I understand that changing
> > > > > > > > state->active_pipes here feels like some unneeded side effect, however having
> > > > > > > > state->active_pipes and bw_state->active_pipes going out of sync doesn't sound
> > > > > > > > very attractive to me either. That is why I don't like this idea of duplication
> > > > > > > > at all - having constant need to sync those state, bw_state, cdclk_state, because
> > > > > > > > they all might have different active_pipes now.
> > > > > > > 
> > > > > > > Having an out of date active_pipes anywhere would be a bug in that
> > > > > > > specific code. Also state->active_pipes is definitely going the way of
> > > > > > > the dodo soon.
> > > > > > > 
> > > > > > > > 
> > > > > > > > > 
> > > > > > > > > > +				intel_calc_active_pipes(state, old_bw_state->active_pipes);
> > > > > > > > > > +			active_pipes_calculated = true;
> > > > > > > > > > +		}
> > > > > > > > > 
> > > > > > > > > I'd do this after the loop so we don't need this extra boolean. As far
> > > > > > > > > as the active_pipes check in intel_crtc_can_enable_sagv(), I think we
> > > > > > > > > can pull it out into intel_compute_sagv_mask() so that we do the check
> > > > > > > > > after computing the mask. And of course change it to use
> > > > > > > > > bw_state->active_pipes instead.
> > > > > > > > 
> > > > > > > > intel_crtc_can_enable_sagv is called per crtc - so can't just pull it out, 
> > > > > > > > will have to have to cycles then - one will compute bw_state->active_pipes,
> > > > > > > > and another pipe_sagv_mask.
> > > > > > > 
> > > > > > > Hmm. Actually I think what we should probably do is keep the
> > > > > > > active_pipes check in intel_can_enable_sagv(). Ie something like this:
> > > > > > > 
> > > > > > > intel_can_enable_sagv(bw_state) {
> > > > > > > 	if (active_pipes && !is_power_of_2(active_pipes))
> > > > > > > 	    	return false;
> > > > > > > 	return sagv_reject != 0;
> > > > > > > }
> > > > > > 
> > > > > > I need active_pipes check here for skl code only, as it disables SAGV for multipipe
> > > > > > scenarios. Adding this here would generalize it for other platforms and we
> > > > > > don't want that for ICL+.
> > > > > 
> > > > > Which is why I said "We can then make the check conditional on pre-icl
> > > > > (or whatever we want) in a later patch". Why in a later patch? Because
> > > > > currently the check is unconditional and it's generally a good idea to
> > > > > limit the number of functional changes per patch to a minimum.
> > > > 
> > > > Moving active_pipes check out of intel_crtc_can_enable_sagv will result
> > > > in wrong SAGV mask calculated.
> > > > 
> > > > i.e if you have 2 pipes,
> > > > 
> > > > for_each_crtc() {
> > > > 	if (crtc_can_sagv())
> > > > 		sagv_reject &= ~pipe;
> > > > 	else
> > > > 		sagv_reject |= pipe;
> > > > }
> > > > 
> > > > will calculate sagv_reject as 0 which is wrong and value will be stored
> > > > in global state.
> > > 
> > > No, it accurately reflects whether each of those pipes is capable of
> > > sagv. The single pipe restriction is an additinal constrain on top of
> > > that. In fact I think adjusting sagv_reject_mask based on this would
> > > result in the wrong value potentially. Consider for example:
> > > 
> > > 1. Enable pipe A + B
> > > 2. sagv_reject would be calculated as 0x3
> > > 3. Disable pipe B
> > > 4. sagv_reject will have pipe B removed, leaving its value at 0x1
> > > 5. No SAGV even though we only have one pipe enabled, which is wrong
> > 
> > Quite good example. I think it means that pipe_sagv_mask should
> > be affected not only by per crtc checks, but also by overall active_pipes
> > state then. I.e ok, I remove active_pipes from intel_crtc_can_enable_sagv,
> > however that active_pipes && !is_power_of_2(active_pipes) check should
> > then also assign pipe_sagv_reject mask to ~0, if it evaluates to true.
> 
> That would clobber the bits of sagv_reject for all crtcs, not just
> the ones part of the state currently. Ie. the very same example
> I gave would still do the wrong thing.

Sounds like we might want to reserve one bit from SAGV mask as "global
SAGV reject" flag, for non-crtc related checks.

I.e if we have active_pipes > 1 for skl, we just set sagv_mask |= 0x80
for instance - that will make your example work and also we will be storing
SAGV result in a single variable.

Stan

> 
> > 
> > I.e then we will always have intel_can_enable_sagv result reflected in
> > sagv mask, which is handy, also intel_can_enable_sagv would still
> > just evaluate sagv mask.
> > 
> > Stan
> > 
> > > 
> > > > I think active_pipes should always affect the SAGV mask
> > > > otherwise we do get really strange situation: you have SAGV mask as 0,
> > > > but you still reject SAGV. So there is no way even then to track what 
> > > > was the previous SAGV state - even if it's 0 it could have been rejected.
> > > > 
> > > > IMO that is quite weird side effect. So removing active_pipes from
> > > > intel_crtc_can_enable_sagv doesn't sound like good idea.
> > > > 
> > > > I think it is now just a bit too much hassle around simple 
> > > > active_pipes_calculated boolean check.
> > > > 
> > > > Stan
> > > > 
> > > > > 
> > > > > > 
> > > > > > In fact that is the only reason I need active pipes here - otherwise I think
> > > > > > it was even your comment that we actually don't need those here at all,
> > > > > > as we just iterate through crtcs in state - pretty clearly remember we discussed
> > > > > > this. Just same way how it's done in intel bw check and other places.
> > > > > > 
> > > > > > Stan
> > > > > > 
> > > > > > > 
> > > > > > > compute_sagv() {
> > > > > > > 	for_each_crtc() {
> > > > > > > 		if (crtc_can_sagv())
> > > > > > > 			sagv_reject &= ~pipe;
> > > > > > > 		else
> > > > > > > 			sagv_reject |= pipe;
> > > > > > > 	}
> > > > > > > 	
> > > > > > > 	active_pipes = calc_active_pipes();
> > > > > > > 
> > > > > > > 	... lock/serialize etc.
> > > > > > > }
> > > > > > > 
> > > > > > > That way we don't have to update sagv_reject at all based on
> > > > > > > active_pipes. I think that even makes more sense since the
> > > > > > > active_pipes check is a global thing and not tied to any specific
> > > > > > > crtc.
> > > > > > > 
> > > > > > > We can then make the check conditional on pre-icl (or whatever we want)
> > > > > > > in a later patch. And finally we can remove it altogether in a separate
> > > > > > > patch, since I don't think we should have to do it on any platform.
> > > > > > > 
> > > > > > > > 
> > > > > > > > > 
> > > > > > > > > We're also going to need to lock_global_state() if bw_state->active_pipes
> > > > > > > > > mask changes.
> > > > > > > > 
> > > > > > > > Ohh.. right.
> > > > > > > > 
> > > > > > > > 
> > > > > > > > Stan
> > > > > > > > 
> > > > > > > > > 
> > > > > > > > > > +
> > > > > > > > > >  		if (intel_crtc_can_enable_sagv(new_crtc_state))
> > > > > > > > > >  			new_bw_state->pipe_sagv_reject &= ~BIT(crtc->pipe);
> > > > > > > > > >  		else
> > > > > > > > > > @@ -5911,11 +5918,9 @@ skl_compute_wm(struct intel_atomic_state *state)
> > > > > > > > > >  	if (ret)
> > > > > > > > > >  		return ret;
> > > > > > > > > >  
> > > > > > > > > > -	if (state->modeset) {
> > > > > > > > > > -		ret = intel_compute_sagv_mask(state);
> > > > > > > > > > -		if (ret)
> > > > > > > > > > -			return ret;
> > > > > > > > > > -	}
> > > > > > > > > > +	ret = intel_compute_sagv_mask(state);
> > > > > > > > > > +	if (ret)
> > > > > > > > > > +		return ret;
> > > > > > > > > 
> > > > > > > > > We also need to remove the state->modeset checks around
> > > > > > > > > sagv_{pre,post}_update().
> > > > > > > > > 
> > > > > > > > > >  
> > > > > > > > > >  	/*
> > > > > > > > > >  	 * skl_compute_ddb() will have adjusted the final watermarks
> > > > > > > > > > -- 
> > > > > > > > > > 2.24.1.485.gad05a3d8e5
> > > > > > > > > 
> > > > > > > > > -- 
> > > > > > > > > Ville Syrjälä
> > > > > > > > > Intel
> > > > > > > 
> > > > > > > -- 
> > > > > > > Ville Syrjälä
> > > > > > > Intel
> > > > > 
> > > > > -- 
> > > > > Ville Syrjälä
> > > > > Intel
> > > 
> > > -- 
> > > Ville Syrjälä
> > > Intel
> 
> -- 
> Ville Syrjälä
> Intel


More information about the Intel-gfx mailing list