Until now, we are initializing and binding to a zwp_dmabuf* instance for each Wayland instance initialization. So for every per-process and per-thread instance that gets initialized. I've been working on reducing the amount of redundant information by deferring this init/binding to when threads enter a gl path that actually calls into dmabuf. I've been trying to think about code paths that might break the present proposal with the default feedback, and for when we introduce per-surface feedback. Please give a review to see if the proposal has merit. Tested with SuperTux Kart.