Re: [PATCH] [RESEND] drm: fb_helper: fix CONFIG_FB dependency

[Arnd Bergmann]

On Wed, Oct 27, 2021 at 1:47 PM Jani Nikula <jani.nikula@xxxxxxxxxxxxxxx> wrote:
> On Thu, 30 Sep 2021, Daniel Vetter <daniel@xxxxxxxx> wrote:
> > On Mon, Sep 27, 2021 at 09:23:45AM -0700, Kees Cook wrote:
> >> On Mon, Sep 27, 2021 at 04:28:02PM +0200, Arnd Bergmann wrote:
> >> > From: Arnd Bergmann <arnd@xxxxxxxx>
> >> >
> >> > With CONFIG_FB=m and CONFIG_DRM=y, we get a link error in the fb helper:
> >> >
> >> > aarch64-linux-ld: drivers/gpu/drm/drm_fb_helper.o: in function `drm_fb_helper_alloc_fbi':
> >> > (.text+0x10cc): undefined reference to `framebuffer_alloc'
> >> >
> >> > Tighten the dependency so it is only allowed in the case that DRM can
> >> > link against FB.
> >> >
> >> > Fixes: f611b1e7624c ("drm: Avoid circular dependencies for CONFIG_FB")
> >> > Link: https://lore.kernel.org/all/20210721152211.2706171-1-arnd@xxxxxxxxxx/
> >> > Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>
> >>
> >> Thanks for fixing this!
> >>
> >> Reviewed-by: Kees Cook <keescook@xxxxxxxxxxxx>
> >
> > Stuffed into drm-misc-next.
>
> The problem is, I don't think the patch is semantically correct.
>
> drm_fb_helper.o is not part of drm.ko, it's part of
> drm_kms_helper.ko. This adds some sort of indirect dependency via DRM
> which might work, maybe by coincidence, maybe not - but it's certainly
> not obvious.

Right, how about this change on top?

--- a/drivers/gpu/drm/Kconfig
+++ b/drivers/gpu/drm/Kconfig
@@ -117,9 +117,8 @@ config DRM_DEBUG_MODESET_LOCK

 config DRM_FBDEV_EMULATION
        bool "Enable legacy fbdev support for your modesetting driver"
-       depends on DRM
-       depends on FB=y || FB=DRM
-       select DRM_KMS_HELPER
+       depends on DRM_KMS_HELPER
+       depends on FB=y || FB=DRM_KMS_HELPER
        select FB_CFB_FILLRECT
        select FB_CFB_COPYAREA
        select FB_CFB_IMAGEBLIT

That would probably make it work for DRM=y, FB=m, DRM_KMS_HELPER=m,
but it needs more randconfig testing, which I can help with.

> The likely culprit is, again, the overuse of select, and in this case
> select DRM_KMS_HELPER. And DRM_KMS_HELPER should depend on FB if
> DRM_FBDEV_EMULATION=y. That's the problem.

This is something we can't easily express in Kconfig, as we can't add the
dependency to a symbol that only gets selected by other drivers, which
is why the dependency has to be in the user-visible symbol,
in this case DRM_FBDEV_EMULATION.

> All of the drm Kconfigs could use an overhaul to be semantically
> correct, but that's a hill nobody wants to die on. Instead we keep
> piling up tweaks to paper over the issues, ad infinitum.

Yes, that is a big issue, though we have similar problems with drivers/media
and net/.

On a related note, I did manage to sort out the backlight dependency issue
(intel_panel.c:(.text+0x2f58): undefined reference to
`backlight_device_register'),
but haven't sent that one again yet, but I can if you like. This one changes
DRM_I915 and all of drivers/video/fbdev from 'select BACKLIGHT_CLASS_DEVICE'
to 'depends on', which I think moves everything into broadly the right
direction.

Let me know if you would like me to send those now, or have a look at the
top 3 patches in [1] if you are interested. This has passed a few
thousand randconfig
builds and should not depend on additional patches.

        Arnd

[1] https://git.kernel.org/pub/scm/linux/kernel/git/arnd/playground.git/log/?h=randconfig-5.16-next