Re: nouveau driver fails to build

[Paul Bolle <pebolle@tiscali.nl>]

On Fri, 2012-11-23 at 18:13 +0100, Helmut Jarausch wrote:
> On 11/23/2012 04:31:19 PM, Paul Bolle wrote:
> > On Fri, 2012-11-23 at 15:42 +0100, Helmut Jarausch wrote:
> > > with 3.7_rc6 and
> > >
> > > CONFIG_ACPI_VIDEO=m
> > > CONFIG_DRM=y
> > > CONFIG_DRM_KMS_HELPER=y
> > > CONFIG_DRM_TTM=y
> > > CONFIG_DRM_NOUVEAU=y
> > > CONFIG_NOUVEAU_DEBUG=5
> > > CONFIG_NOUVEAU_DEBUG_DEFAULT=3
> > > CONFIG_DRM_NOUVEAU_BACKLIGHT=y
> > >
> > > with/ without
> > > CONFIG_DRM_LOAD_EDID_FIRMWARE=y
> > >
> > >
> > > The kernel fails to build with
> > > drivers/built-in.o: In function `nouveau_acpi_edid':
> > > (.text+0x12bf21): undefined reference to `acpi_video_get_edid'
> > >
> > >
> > > What am I missing?
> > 
> > I'd guess it should build (at least without this error) with
> > CONFIG_ACPI_VIDEO=y. Is that correct?
> 
> Yes, but this is very unnatural.
> 
> To get CONFIG_ACPI_VIDEO=y  I need to have VIDEO_OUTPUT_CONTROL=y
> for which I have to select the Intel GMA500 Stub Driver although I don't
> have such a device and therefore don't need that driver.
> Furthermore, I had to chase through quite some flags to find this.
> 
> If it's not a bug, it's surprising, at least.

0) It puzzles me too.

1) The problem is in Nouveau's Kconfig:
    config DRM_NOUVEAU
        tristate "Nouveau (nVidia) cards"
        [...]
        select ACPI_VIDEO if ACPI && X86 && BACKLIGHT_CLASS_DEVICE && VIDEO_OUTPUT_CONTROL && INPUT
        [...]

and its Makefile:
    # other random bits
    [...]
    nouveau-$(CONFIG_ACPI) += nouveau_acpi.o
    [...]

    obj-$(CONFIG_DRM_NOUVEAU)+= nouveau.o

2) Now, with DRM_NOUVEAU=y and ACPI_VIDEO=m nouveau_acpi.o will be made
builtin (which fails with the error Helmut ran into). The puzzling part
is that ACPI_VIDEO will only be selected (ie, made into a 'y') if all
dependencies of that select statement are 'y' too. (Note that these
dependencies are identical to the dependencies of ACPI_VIDEO's config
entry. That can be no coincidence.)

3) But I would expect ACPI_VIDEO to be selected if the dependencies of
that select statement are either 'y' or 'm'. And then Helmut's .config
should trigger some warning about selecting a symbol with unmet
dependencies, or whatever, shouldn't it? (Because some of ACPI_VIDEO's
dependencies are still 'm'.)

4) I'm CC-ing the Kconfig people to tell us whether the current behavior
is as designed or a bug. If it's not a bug the DRM_NOUVEAU config entry
might need to be changed.


Paul Bolle