[elrepo] Problem with CUDA since 331.67.elrepo

Phil Perry phil at elrepo.org
Sat Apr 26 07:41:41 EDT 2014


Hi Michael,

On 25/04/14 00:31, Michael Lampe wrote:
> They now have a second kernel module 'nvidia-uvm' that wants to be built
> and packaged too. Affected are both older CUDA versions and the new 6.0.
>

Yes, I noticed that module added a while back, but have yet to package 
it as I'm not a CUDA user and had no idea of it's intended usage nor 
facility to test. Now we have some real users here I'm happy to oblige :-)

> While that's easy to fix, there is an annoying little problem. There
> comes a new device file '/dev/nvidia-uvm' with the new driver module.
> And it picks an arbitrary unused major no. at load time. So you cannot
> have it loaded by modprobe automatically when the device node is
> accessed, because you cannot create the device node before the module is
> loaded, and you also cannot use udev in any way because this module
> creates no data in sysfs and therfore creates no udev events.
>
> But the usage pattern is that CUDA programs will access
> '/dev/nvidia-uvm' and expect everything to be already in place.
>
> The only solution I found is to piggyback the new module + device
> creation on the old one:
>
> $ cat /etc/modprobe.d/nvidia
> options nvidia NVreg_ModifyDeviceFiles=0
> install nvidia /sbin/modprobe --ignore-install nvidia; /sbin/modprobe
> nvidia-uvm; /bin/mknod -m 0600 /dev/nvidia-uvm c `/bin/grep nvidia-uvm
> /proc/devices | /usr/bin/cut -d' ' -f1` 0; /sbin/pam_console_apply
> /dev/nvidia-uvm
>
> This works because '/proc/devices' reveals what major the new module
> picked. Permissions are set and (automatically) maintained with the same
> logic that applies to the other /dev/nvidia* files.
>
> What do you think? If nobody comes up with a better idea, I'll post the
> complete patch.
>

A couple of questions here, to help me understand the options available 
to us.

Is this anything nvidia-modprobe can handle? Should we be packaging this 
file (I don't believe we are are present)?

I'd also like to review the rpmfusion package to see if/how the Fedora 
folks handle this.

I will need you guys, the folks that actually use this stuff, to guide 
me as to what you require and what works.

Thanks,

Phil




More information about the elrepo mailing list