I am a bit concerned about the performance implication of accessing the thread wakeup through a lock in the call queue, but for now I think it's reasonable to address the buildbot failure with a locking solution while we try to find a better one.

I'm not certain if we'll be able to find one that correctly addresses the failures with zero additional locking, but I think we may be able to reduce the number of times we use the lock compared to the implementation in GH-19760. I'll spend some time exploring that as I find the time to, and report back with any significant findings.