For cpu syncs, with async_cb, fence_retire gets called inside render worker submit_fence. Sync thread of proxy will catch it right away and retire the fence before the fence gets added to the timeline. Without this CL, the guest cpu syncs will always wait until timeout. Any following sync points added to the global fencing timeline at virtio-gpu device process will be blocked. As long as there's a fence on the timeline must be waited to signal (e.g. wsi purpose), a deadlock occurs.
Similar fix(30bb18d3) has landed previously for virgl and the sync threads of its sub-ctx. This fix handles the similar in between proxy and server.
TEST: ANGLE on Venus works normally with