> acquiring locks before fork in the thread doing the forking and releasing them afterwards is always the safe thing to do.

It's also an easy way to cause a deadlock:

- register_at_fork() et.al. will cause per-logger locks to be acquired before the global lock (this isn't immediately obvious from the code)

If a thread were to grab its logging lock before the global lock then it would deadlock.  I'm guessing this isn't allowed - however I didn't see any comments to this effect?

Can I suggest documenting this, and also merging the two callbacks so that the ordering of these two acquires is made explicit.

- the per-logger locks are acquired in a random order

If a thread were to acquire two per-logger locks in a different order then things would deadlock.