[Intel-gfx] [PATCH] drm/i915/psr: Split sel fetch plane configuration into arm and noarm

[Luca Coelho]

On Fri, 2023-01-27 at 11:33 -0800, Lucas De Marchi wrote:
> On Fri, Jan 27, 2023 at 07:12:29PM +0200, Luca Coelho wrote:
> > On Fri, 2023-01-27 at 16:37 +0200, Jani Nikula wrote:
> > > On Fri, 27 Jan 2023, Tvrtko Ursulin <tvrtko.ursulin at linux.intel.com> wrote:
> > > > On 26/01/2023 16:05, Jani Nikula wrote:
> > > > > On Thu, 26 Jan 2023, Luca Coelho <luca at coelho.fi> wrote:
> > > > > > On Thu, 2023-01-26 at 14:11 +0200, Luca Coelho wrote:
> > > > > > > On Thu, 2023-01-26 at 14:00 +0200, Jani Nikula wrote:
> > > > > > > > On Thu, 26 Jan 2023, Luca Coelho <luca at coelho.fi> wrote:
> > > > > > > > > On Wed, 2023-01-25 at 12:44 +0200, Jouni Högander wrote:
> > > > > > > > > > > diff --git a/drivers/gpu/drm/i915/display/intel_psr.c b/drivers/gpu/drm/i915/display/intel_psr.c
> > > > > > > > > > > index 7d4a15a283a0..63b79c611932 100644
> > > > > > > > > > > --- a/drivers/gpu/drm/i915/display/intel_psr.c
> > > > > > > > > > > +++ b/drivers/gpu/drm/i915/display/intel_psr.c
> > > > > > > > > > > @@ -1559,7 +1559,26 @@ void intel_psr2_disable_plane_sel_fetch(struct intel_plane *plane,
> > > > > > > > > > >   	intel_de_write_fw(dev_priv, PLANE_SEL_FETCH_CTL(pipe, plane->id), 0);
> > > > > > > > > > >   }
> > > > > > > > > > > 
> > > > > > > > > > > -void intel_psr2_program_plane_sel_fetch(struct intel_plane *plane,
> > > > > > > > > > > +void intel_psr2_program_plane_sel_fetch_arm(struct intel_plane *plane,
> > > > > > > > > > > +					const struct intel_crtc_state *crtc_state,
> > > > > > > > > > > +					const struct intel_plane_state *plane_state,
> > > > > > > > > > > +					int color_plane)
> > > > > > > > > > > +{
> > > > > > > > > > > +	struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
> > > > > > > > > 
> > > > > > > > > Should you use i915 instead of dev_priv? I've heard and read elsewhere
> > > > > > > > > that this is generally a desired change.  Much easier to use always the
> > > > > > > > > same local name for this kind of thing.  Though this file is already
> > > > > > > > > interspersed with both versions...
> > > > > > > > 
> > > > > > > > Basically the only reason to use dev_priv for new code is to deal with
> > > > > > > > some register macros that still have implicit dev_priv in
> > > > > > > > them. Otherwise, i915 should be used, and when convenient, dev_priv
> > > > > > > > should be converted to i915 while touching the code anyway (in a
> > > > > > > > separate patch, but while you're there).
> > > > > > > 
> > > > > > > Thanks for the clarification! In this case we're not using any of the
> > > > > > > macros, AFAICT, so I guess it's better to go with i915 already? And I
> > > > > > > think it should even be in this same patch, since it's a new function
> > > > > > > anyway.
> > > > > > > 
> > > > > > > 
> > > > > > > > The implicit dev_priv dependencies in the register macros are a bit
> > > > > > > > annoying to fix, and it's been going slow. In retrospect maybe the right
> > > > > > > > thing would have been to just sed the parameter to all of them
> > > > > > > > everywhere and be done with it for good. Not too late now, I guess, and
> > > > > > > > I'd take the patches in a heartbeat if someone were to step up and do
> > > > > > > > it.
> > > > > > > 
> > > > > > > I see that there is a boatload of register macros using it... I won't
> > > > > > > promise, but I think it would be a good exercise for a n00b like me to
> > > > > > > make this patch, though I already foresee another boatload of conflicts
> > > > > > > with the internal trees and everything...
> > > > > > 
> > > > > > There were actually 10 boatloads of places to change:
> > > > > > 
> > > > > >   187 files changed, 12104 insertions(+), 12104 deletions(-)
> > > > > > 
> > > > > > ...but it _does_ compile. 😄
> > > > > > 
> > > > > > Do you think this is fine? Lots of shuffle, but if you think it's okay,
> > > > > > I can send the patch out now.
> > > > > 
> > > > > Heh, I said I'd take patchES, not everything together! ;)
> > > > > 
> > > > > Rodrigo, Tvrtko, Joonas, thoughts?
> > > > 
> > > > IMO if the elimination of implicit dev_priv is not included then I am
> > > > not sure the churn is worth the effort.
> > > > 
> > > > I think one trap is that it is easy to assume solving those conflicts is
> > > > easy because there is a script, somewhere, whatever, but one needs to be
> > > > careful with assuming a random person hitting a merge conflict will
> > > > realize there is a script, know where to find it, and know how to use it
> > > > against a state where conflict markers are sitting in their local tree.
> > > > That's a lot of assumed knowledge which my experience tells me is not
> > > > universally there.
> > > > 
> > > > Having said all that, I looked at the occurrence histogram for the
> > > > proposed churn and gut feel says conflicts wouldn't even be that bad
> > > > since they seem heavily localized in a handful of files plus the display
> > > > subdir.
> > > > 
> > > > Plus it is upstream.. so we are allowed not to care too much about
> > > > backporting woes. I would still hope implicit dev_priv, albeit
> > > > orthogonal, would be coming somewhat together with the rename. For that
> > > > warm fuzzy feeling that the churn was really really worth it.
> > > 
> > > I was mostly talking about the implicit dev_priv removal. It's somewhat
> > > easy, because you can always assume dev_priv is around when the macros
> > > in question are used.
> > > 
> > > The above is a dependency to any renames. I don't think the renames are
> > > as important as removing the implicit dev_priv, and the renames are
> > > easier to handle piecemeal, say a file at a time or something.
> > 
> > I'm trying to write a semantic patch to convert this stuff.  But
> > coccinelle is problematic when it comes to macros, so it turned out not
> > to be as trivial as I though.
> 
> I think that the definition in the header is easier to do manually and
> let coccinelle change only the users. I started this and it seems to be
> going the right direction:
> 
> 2 prerequisite commits:
> https://git.kernel.org/pub/scm/linux/kernel/git/demarchi/linux.git/log/?h=tip-drm-intel-dev-priv
> 
> $ cat /tmp/a.cocci
> virtual patch
> 
> @@
> expression e;
> @@
> - DPLL(e)
> + DPLL(dev_priv, e)
> 
> @@
> expression e;
> @@
> - DPLL_MD(e)
> + DPLL_MD(dev_priv, e)
> 
> @@
> expression e1, e2;
> @@
> - PALETTE(e1, e2)
> + PALETTE(dev_priv, e1, e2)
> 
> ... simply continuing with the same pattern for the other macros
> I *think* would produce a good result. I slightly tested it with
> `make coccicheck MODE=patch COCCI=/tmp/a.cocci  M=./drivers/gpu/drm/i915`

Yes, this probably works, but it is more "seddy" than semantic.  My
goal was to find rules that will change all these occurrences without
having to specify each different patch individually. ;)


> Then if we change the macro in i915_reg.h we could remove all the
> implicit deps. Wether we should pass dev_priv or mmio_base I think will
> vary from macro to macro.  The rename s/dev_priv/i915/ being the cherry
> on top.

Yes, the changes are a bit independent, but we'll have to do one on top
of the other, of course.

--
Cheers,
Luca.