For x86-64, arch/x86/lib/Makefile has:
        lib-y += csum-partial_64.o csum-copy_64.o csum-wrappers_64.o
so in particular ip_compute_csum() (from csum-partial_64.c) gets put in

libs, so even though it is exported, it gets thrown away at link time if

there are no non-modular users.
However, if someone has an insane config with CONFIG_NET=n and

CONFIG_VIDEO_BT848=m, then ip_compute_csum() gets discarded but the

modular code in drivers/media/video/bt8xx/bttv-cards.c calls it and...
        ERROR: "ip_compute_csum" [drivers/media/video/bt8xx/bttv.ko] undefined!
I guess one fix is just to put at least csum-partial_64.c in obj-y

instead; it does seem broken to have files with EXPORT_SYMBOL() be

lib-y, since it is really an invitation for this problem.  Other ideas

would be for bttv to depend on or select CONFIG_NET.
[It's a long story why I noticed this problem, but maybe we want to fix

it to save people doing randconfig test builds or something like that.

Or maybe we don't really care, since the class of real systems with

video capture cards but no networking is probably quite small]
 - R.

--

This isn't the first time csum-partial_64.c is involved in such a kind

of breakage (and we constantly have this kind of bugs with various

exports from lib-y), and EXPORT_SYMBOL's in lib-y are simply wrong and

always bugs waiting to happen.
cu

Adrian
-- 
       "Is there not promise of rain?" Ling Tan asked suddenly out

        of the darkness. There had been need of rain for many days.

       "Only a promise," Lao Er said.

                                       Pearl S. Buck - Dragon Seed
--

well, the TCP/IP checksumming functions are in lib portions on every

architecture i checked, and the exports are handled by the networking

code. And the symbol used is named ip_compute_csum(). So i think the

correct solution is to make CONFIG_VIDEO_BT848 depend on CONFIG_NET ...
especially on embedded there's an advantage in linking out unused core

kernel functions. But even if we did this change on x86, other

architectures might (correctly) complain about x86 going it alone. x86

gets the most use by far, so if we start introducing build assymetries

like this it will just mess up the build on other architectures.
	Ingo

--