Bug#764639: libgl1-nvidia-glx: failst to upgrade or install with KMS (nouveau) active

[Andreas Beckmann]

Control: fixed -1 361.45.18-1

On 2017-02-11 00:06, Andreas Beckmann wrote:
> Problem is that the 340xx driver (didn't check 304xx) does not
> unregister the proc entries if it aborts loading (due to nouveau being
> loaded already).

The missing nv_unregister_procfs() call was added upstream in the 361
series.

The same problem seems to exist in 304xx. I've ported my patch to that
legacy series, too, but I have no way to test it properly.

Luca, perhaps you could try the following (even if you have no 304xx
requiring hardware, as long as the test machine has a nvidia gpu):
* load nouveau module
* load 304xx module (should fail, since another module has already
grabbed the device)

I don't know if the relevant code path gets executed if no supported
hardware is available ... if it is I would expect the 304xx module
(without patch) to leave a stale /proc/driver/nvidia (requires reboot to
get rid of it again), while the patched version leaves /proc clean.
It's also possible that the problem is not (easily) reproducible on
304xx, since the initialization of the procfs entries seems to come much
later.


Andreas