New installer vs. 2.4.20: nvidia.o (graphics) = nvidia.o (nForce IDE)

[stuartlevy]

Puzzled why our new Linux machine with NV30 (Quadro 2000 FX?),
RH 7.2, XF86 4.2.0, NVidia 4349 from new installer, didn't work:
claiming it couldn't initialize the
kernel driver even though lsmod listed
"nvidia" as loaded!

Finally noticed: I'd built the 2.4.20 kernel with
CONFIG_BLK_DEV_NFORCE, the
IDE driver for the NVidia motherboard chipset.
Our motherboard doesn't have
one, but I figured it was harmless to
enable it.

Not so: the name of that module is, yes,
"nvidia.o", the same name as the graphics
driver!! So when the kernel autoloaded
"modprobe nvidia", it got the wrong module.

So Nvidia installer designers, could you
maybe change the module name?
nvidiagfx.o, say? You might argue that it
wouldn't make sense
to plug an NVidia graphics card into an NForce motherboard,
but others (including software distro builders)
might do just what I did and
cause trouble.

[bwkaz]

Or get the kernel developers to rename the IDE driver to nforceide instead of nvidia, since that's what it is.

Or, you could insmod /full/path/to/the/right/nvidia.o instead.

[stuartlevy]

Given that the kernel developers generally
consider binary-only drivers as non-Kosher,
I wouldn't have much hope in asking them to
change one of the stock module's names to
accommodate a proprietary driver.

It's true, once an individual user knows what
the problem is, it would be easy to work around --
by modprobe'ing the full path,
or manually renaming the module file.
But the point is that a name collision is
hidden breakage: most of the difficulty
was in discovering why this setup, which
normally worked nicely, was failing this time
without any meaningful diagnostics appearing anywhere.

We need to make sure this doesn't happen!
So I ask the NVidia developers to help.

It might be a good idea (and the core
kernel people might like this) to make modprobe detect ambiguities:
if you "modprobe nvidia" it could have warned in the syslog that
it had found several nvidia.o's and chosen one.
This kind of thing will surely happen again in other contexts.

[ripperda]

which kernel version is this? the standard 2.4.20 from kernel.org doesn't have this (that I can tell)

[stuartlevy]

Originally posted by ripperda
which kernel version is this? the standard 2.4.20 from kernel.org doesn't have this (that I can tell)

Ow -- you're right, I was withholding information.

Looking back, I'd assembled it from
stock 2.4.20 plus the 2.4.21-pre3 patch.

I just checked the latest kernel.org prerelease
of 2.4.21 (2.4.21-rc2). It no longer includes
any CONFIG_BLK_DEV_NVIDIA option --
the nForce IDE driver that was present in pre3
seems to be gone in rc2. So the nvidia.o collision is
probably a non-issue for everyone except
ham-handed tinkerers like me, who deserve what we get.
If somebody adds it back to the main kernel
line they'll probably give it a different name.

[stuartlevy]

Checking a bit more carefully, the 2.4.21-rc2 does still have an NVidia nForce
IDE tuning driver. But it's now part of another module:
CONFIG_BLK_DEV_AMD74XX, or amd74xx.o.
"make *config" lists it as "AMD and nVidia IDE support". So all's well, as far as I can tell.