Store the driver top-level DSOs to /run/opengl-driver

[picnoir]

Idea of @soupglasses

We would still modify the rpaths to point to /var/lib

Pros:

• Can use Nix binaries without wrapping them first.
• Fixes the wrapper startup overhead.

Cons:

• Need something to keep /run/opengl-driver in sync.
  • We can do that using the package managers hooks. That means we have to distribute nix-gl-host through the host package manager though.
  • We can do that using a nix-distributed daemon epolling the top-level DSOs.

[knedlsepp]

I've been using a very similar solution quite successfully on RHEL for some CUDA programs. Instead of a daemon epolling the DSOs I'm creating a systemd .path unit like so:

[Path]
Unit=copy-and-patchelf-libcuda-and-libnvidia.service
PathModified=/usr/lib64/libcuda.so

[Install]
WantedBy=multi-user.target

And in the copy-and-patchelf-libcuda-and-libnvidia.service I copy and patchelf /usr/lib64/libcuda.so* /usr/lib64/libnvidia*. Sadly PathModifiedGlob doesn't exist (yet) so that approach only works if libcuda.so is changed.