The Right Way, actually, is not to use host GL library, but to put Mesa + nvidia's library inside the runtime, and rely on the kernel's userspace API...