Actually, there is a problem in Lock.acquire and RLock.acquire. If a signal occurs and signal handling returns successfully, acquiring the lock will be retried without decrementing the timeout first. Therefore, we may end up waiting longer than the user wanted.

I'm not sure how to tackle that: either we accept that an incoming signal will make the wait longer, or we fix it by properly decrementing the timeout (which will complicate things a bit, especially for cross-platform time querying - but see issue9079 which might help us).