Added a flag to allow for at least one run -- I know nothing of non-Linux clock resolution.  That should handle that.

As for the thread safety of the socket timeouts, yeah, that was why I didn't do that initially, I assumed the suggestion to take that approach took the risk into account; you'll know far more about potential impact than I will.

Since this is at a higher abstraction than socket primitives, another option would be to track remaining time in thread local data so that we don't mutate the timeout on the object (which I don't really like doing anyway). 

Thoughts on approach before I put it together?