I don't think we necessarily want to apply this large of a hammer.

In particular, I don't think we want to clobber any parts of the mask that aren't SIGCHLD on those systems that actually try to propagate updates to other threads with sigprocmask(). On FreeBSD, we're potentially restoring a completely unrelated thread's signal mask to the current thread.

Actually, after writing out that last bit, that does mean that FreeBSD should be doing the mask manipulation on every call instead of just on the first/last.

@dim in your opinion, should concurrent here be volatile, or is the fact that it's static sufficient to deter optimization?

I think static or not doesn't matter for whether the compiler will decide to reload the variable even though it (from its perspective) could not have changed. If other threads are updating the same variable concurrently, you should ideally use atomic operations, but volatile will also work in this case.

It was atomic in a previous version of the patch, but the spinlock (which is needed for other reasons) makes that unnecessary, I think. I'll add volatile.