For this use case a major one would be better models for carved up shared memory with safe/secure mappings in and out of specialized hardware like the gpu. Android uses binder for this and there are a good number of practical pains with it being shoved into that shape. Some other teams at Google doing similar stuff at least briefly had a path with another kernel module to expose a lot more and it apparently enabled them to fix a lot of problems with contention and so on. So it’s possible to solve this kind of stuff, just painful to be missing the primitives.