Manually resetting this is not safe when exceptions happen. Since it can't be done in a finally block this should use RAII to ensure that value gets reset to false.

Same below.

Do you suggest setting it to false in the destructor, then?

I refer to calling spinning.exchange(true) before in this function. If an exception happens afterwards it doesn't get set back to false. This should use RAII (like a scoped mutex lock) to ensure setting the value back in any case.

Got it.

I was trying to avoid adding mutexes to Executor where they're not needed. But mutex and lock_guard are the most suited to RAII-style management of this state. Maybe this would work with try_lock instead of lock so that we throw an exception if the Executor is already spinning instead of blocking.

@wjwwood, since you implemented about half of this PR, any thoughts?

Yeah I'll send another pr asap.

Yeah, my first pass at the strategy I suggested was not successful (deadlock) and it makes the code a lot more confusing.

Sorry, I just realized you were expecting me to do this 😄. I totally misread it, but anyways I just pushed my fix. I did it by implementing what's called a scope exit or scope guard. Boost has a scope exit and so does the D language. There's some interesting discussion about including something like this in the c++ std library at some point. Anyways this should work us and might be useful else where too.

With this implemented should we remove the redundant setting to false at the end of the scope?

That's a fair point, I'll push another commit.

f1e7ea5