Volume setup fails across device boundaries

[mvollrath]

When your Docker graph is not mounted on the same device as /usr/bin etc., you see these sorts of errors during nvidia-docker volume setup:

Error: link /usr/bin/nvidia-cuda-mps-control /mnt/docker/volumes/nvidia_driver_352.79/_data/bin/nvidia-cuda-mps-control: invalid cross-device link

This is caused by hard-linking the binaries here:

nvidia-docker/tools/src/nvidia/volumes.go

Line 179 in e7b7922

if err := os.Link(f, l); err != nil {

[3XX0]

It has been discussed in #38 and #34
Copying driver files is not recommended, the overhead is non negligible (100+MB) and it can potentially confuse the NVIDIA uninstaller thus interfering with drivers upgrades.

If we really want to do it, it should be on the volume setup only. Also keep in mind that the nvidia package can be used outside of the nvidia-docker project, the behavior of Create needs to be defined (does it create links or copies?).

In any case, relying on the plugin to manage the volumes is still preferred.

[mvollrath]

Since the plugin doesn't support a Quadro K600, that's not an option for me right now. #40

I'm integrating nvidia-docker with a system that boots from a read-only image, so the entire root filesystem is overlayfs with the NV driver already installed. Running the Docker graph on overlay would quickly exhaust it, so a fresh logical volume is prepared at boot. It looks like I'll have to keep running with my fork until either plugin compatibility is fixed or I can refactor this copy fallback to only affect volume setup.