On Wed, May 22, 2013 at 06:22:12PM -0400, Michael R. Hines wrote: > Hi, > > We run nvidia devices inside libvirt-managed LXC containers. > > It used to be that simply doing: > > $ echo 'c 195:* rwm' > /sys/fs/cgroup/devices/libvirt/lxc > > Then, after booting the container, we would do: > > $ mknod -m 666 /dev/nvidia0 c 195 0 > > .... would be good enough to run our CUDA applications. > > But, according to: > > $ cat src/lxc/lxc_container.c > > The CAP_MKNOD capability is being dropped and only a specific > set of devices is being created before booting the container. > > Is there any reason why this is not per-device configurable? With recent libvirt you can pass through arbitrary block and character devices explicitly, using the following XML: http://libvirt.org/formatdomain.html#elementsHostDevCaps As such there is never any need to change cgroups or use mknod as you describe. Daniel