Hi Maxime, On Wed Sep 2, 2026 at 2:36 PM CEST, Maxime Ripard wrote: > On Tue, Sep 01, 2026 at 04:30:40PM +0200, Luca Ceresoli wrote: >> On Tue Sep 1, 2026 at 2:40 PM CEST, Luca Ceresoli wrote: >> >> [...] >> >> >>> >> > And we'd essentially move drm_panel_bridge into drm_panel.c, and >> >>> >> > make it >> >>> >> > private. >> >>> >> >> >>> >> Yes in theory, but the panel_bridge code uses other parts of the >> >>> >> drm_kms_helper module: drm_atomic_helper and drm_probe_helper, maybe >> >>> >> more, >> >>> >> so we'd have to move them into the drm module too. >> >>> > >> >>> > Ah, right. What would happen if we were doing it the other way around >> >>> > then? Move drm_panel out of the main drm module? >> >>> >> >>> Into the drm_kms_helper module? >> >>> >> >>> I had a look and did some experiments and I found at least one user in >> >>> the drm >> >>> module calling a drm_panel API, and guess who: >> >>> >> >>> drm_of_find_panel_or_bridge() (in drm_of.c, drm module) >> >>> -> calls of_drm_find_panel (in drm_panel.c, drm_kms_helper module) >> >>> >> >>> Based on our discussion after patch 3, I'm not sure >> >>> drm_of_find_panel_or_bridge() will disappear soon. If it doesn't, I guess >> >>> we can try to move drm_of_find_panel_or_bridge() into bridge/panel.c >> >>> which >> >>> is in the drm_kms_helper module (and from drm_of.h info >> >>> drm_bridge.h?). That however might trigger build failures for drivers >> >>> which >> >>> currently don't select DRM_KMS_HELPER and which would have to select it. >> >>> >> >>> I'll give it a try, and if I see major drawbacks I will get back to >> >>> moving >> >>> all the drm_kms_helper code into the main drm module. >> >> >> >> No, I meant into its own module. >> > >> > Looks like a good idea indeed, making the design more modular and >> > dependencies cleaner. >> > >> >> Do we have any dependency from the main >> >> drm module into drm_panel? >> > >> > As far as I can see there is only the one mentioned above: >> > >> > drm_of_find_panel_or_bridge() (in drm_of.c, drm module) >> > -> calls of_drm_find_panel (in drm_panel.c, would-be the new >> > drm_panel module) >> > >> > And while drm_of_find_panel_or_bridge() is possibly going to disappear at >> > the end of my series, moving drm_panel.c to its own module would make the >> > series non-build-bisectable, which would be very annoying. >> > >> > However there seem to be no calls from the drm module to >> > drm_of_find_panel_or_bridge(), so we could move it out as well. What about >> > a series doing, in this order: >> > >> > - move drm_of_find_panel_or_bridge() from the drm module to the >> > drm_kms_helper module >> > - move drm_panel.c to a new drm_panel module >> > - Main change: embed a drm_bridge into every panel >> > - convert drivers to stop using the panel_bridge, hopefully removing all >> > calls to drm_of_find_panel_or_bridge() >> > - remove drm_of_find_panel_or_bridge() >> >> Ah, no, that won't work. There would be a circular module dependency loop >> later on when we embed a drm_bridge into evern drm_panel: indeed at that >> point drm_panel will use the atomic and probe helpers to implement the >> embedded drm_bridge, resulting in: >> >> * the panel_bridge code in bridge/panel.c [drm_kms_helper module] >> already depends on the drm_panel.c code (it manipulates a drm_panel, OK) >> * additionally, the drm_panel.c code, in order to create a drm_bridge, >> will depend on the helpers in drm_atomic_helper and drm_probe_helper >> code [drm_kms_helper module] >> >> The loop is only between kernel modules (.ko), not in actual code. So I >> think this revised plan should work (the 2nd bullet is key): >> >> - move drm_of_find_panel_or_bridge() from the drm module to >> bridge/panel.c [currently drm_kms_helper module] >> - move bridge/panel.o to a new drm_panel_bridge module (NEW) >> - move drm_panel.c to a new drm_panel module > > Looks good on principle, but iirc the starting point of that discussion > was to move bridge panel into the new panel module,
That was the starting point, but working on it I realized it was very tricky. So I tried instead copying the relevant bridge/panel.c code into drm_panel.c, but with more suitable function names, and the result is way cleaner. This changed the steps, but not the end result: remove drm_of_find_panel_or_bridge() and bridge/panel.c entirely. > so I guess we could: > > 1) Move drm_of_find_panel_or_bridge() to drm_panel.c. Both are still in > the drm module at this point, so it should be ok. > 2) Create a new panel module, with a dependency on bridge > 3) Move the bridge/panel.c code into the new panel module to create the > bridge at the same time we create the panel. > > Would that work? Maybe, but as above, my plan has changed a bit. My opinion is that for this tricky task we should focus more on the end result (embed a bridge into every panel, remove drm_of_find_panel_or_bridge() and bridge/panel.c) than on finding the most optimal path to get there. So I'm aiming at a path that is build-bisectable, clean and understandable. What do you think about this principle? I'm sending a RFC v2 soon, so we can discuss on a specific proposal. Luca -- Luca Ceresoli, Bootlin Embedded Linux and Kernel engineering https://bootlin.com
