[PATCH 1/2] drm/nouveau: Add DRM_IOCTL_NOUVEAU_GET_ZCULL_INFO

[M Henning]

On Thu, Mar 27, 2025 at 8:56 AM Danilo Krummrich <dakr at kernel.org> wrote:
>
> On Tue, Mar 25, 2025 at 07:40:56PM -0400, M Henning wrote:
> > Okay, that sounds reasonable since I don't expect this to change very quickly.
> >
> > Since I don't fully understand, is the suggestion here to:
> > 1) add the interface as a function on nvkm_gr using the nvkm_gr_func
> > vtable and store the actual data on r535_gr
> > or
> > 2) add the interface to NVIF (which IF?) and store the actual data on nvkm_gr
> > ?
>
> I think we want both.
>
> 1) I think the suggestion was to store the data directly in nvkm_gr, however the
>    structure is indeed specific to r535, so I think, unfortunately, we need the
>    vtable and store that data in r535_gr.

Well, NV2080_CTRL_GR_GET_ZCULL_INFO_PARAMS is r535-specific, but we
need to convert it into a common structure and combine it with info
from NV0080_CTRL_FIFO_GET_ENGINE_CONTEXT_PROPERTIES at some point, so
I think it makes sense to do that conversion+combination before
storing it on any structure. In that case, maybe we store the
structure on nvkm_gr directly during r535_gr_oneinit and then the call
to get the info only goes through NVIF?


> 2) Yes, this should be passed through nvif. Unfortunately, I think it would need
>    to be a whole new one (similar to the fifo one).
>
> Maybe Ben can provide you some additional pointers one this? Mayber he can
> suggest a shortcut, since he has patches queued to simplify the whole interface.
>
> >
> > (Sorry, I don't understand how these layers are intended to fit together.))
> >
> > On Thu, Mar 20, 2025 at 4:02 PM Danilo Krummrich <dakr at kernel.org> wrote:
> > >
> > > On Fri, Mar 21, 2025 at 05:57:55AM +1000, Ben Skeggs wrote:
> > > > On 21/3/25 04:18, Danilo Krummrich wrote:
> > > >
> > > > > Hi Mel,
> > > > >
> > > > > On Wed, Mar 12, 2025 at 05:36:14PM -0400, Mel Henning wrote:
> > > >
> > > > > > @@ -72,6 +75,9 @@ struct nvkm_device {
> > > > > >                   bool armed;
> > > > > >                   bool legacy_done;
> > > > > >           } intr;
> > > > > > +
> > > > > > + bool has_zcull_info;
> > > > > > + struct drm_nouveau_get_zcull_info zcull_info;
> > > > > This is bypassing the nvif layer entirely. I think you should store the contents
> > > > > of struct NV2080_CTRL_GR_GET_ZCULL_INFO_PARAMS in struct r535_gr and have an
> > > > > nvif interface to access the data.
> > > >
> > > > I agree here, though nvkm_gr would be a better choice for a couple of
> > > > reasons, not least that it's possible (and should be reasonably trivial) to
> > > > support this on earlier GPUs - should someone desire to at a later point.
> > >
> > > I agree, if the interface is stable enough -- I don't know whether this is prone
> > > to change or not.