Should condition_variable.notify_all be covered by mutex lock?

Question

I have implemented a class that allows me to synchronize threads with a condition variable. I have found conflicting information as to whether the notify_all should be done

Answer 1

There is no risk releasing the lock if the mutex was held in some interval between the state of the condition test changing and thr notification.

{
    std::lock_guard lk(_address_mutex);
    _address = address;
}
_cv.notify_all();

here, the mutex was unlocked after _address was changed. So no risk.

If we modify _address to be atomic, naively this looks correct:

{
    std::lock_guard lk(_address_mutex);
}
_address = address;
_cv.notify_all();

but it is not; here, the mutex is released for the entire period between modifying the condition test and notification,

_address = address;
{
    std::lock_guard lk(_address_mutex);
}
_cv.notify_all();

The above, however, becomes correct again (if more than a little strange).

---

The risk is that the condition test will be evaluated with the mutex active (as false), then changed, then notification sent, then the waiting thread waits on a notification and releases the mutex.

waiting|signalling
lock
test
        test changed
        notification
listen+unlock

the above is an example of a missed notification.

So long as we have the mutex held anywhere after the test change and before the notification it cannot happen.