Forgive me for a short rant, but this sort of thing is one reason I think other kernels should look at OS X's virtual memory interface for inspiration. There, you can always share any existing mapped page with another process, regardless of how it was allocated. Usually useful because it reduces the need to redo the way you allocate some (regular) memory just because you want to use it a certain way (sharing), but it works with driver-mapped pages too.

Sharing a normal memory mapping is not necessarily what is required here though. What you want is to share the same GPU-side memory buffer between two processes. It may not even be mappable to cpu side memory.

If it's mapped into the CPU address space, then it can be shared. If not, then there is a copy going on and you may as well just do it in the non-sandboxed process, no?

There is no copy. You have a reference to a GPU side buffer, and you want to pass a reference to that buffer to another process, where it will continue to use the GPU to access it.