postgres deadlock_timeout - how does it work exactly

Here is my understanding of how deadlock_timeout behaves in postgres - am I correct?

• deadlock_timeout is a setting on transaction level, it can be changed via set local, so each transaction runs the 'deadlock detection logic' by itself and there is no such logic on a level higher than on transaction level
• When the transaction starts waiting for a lock, it waits deadlock_timeout seconds, if it is then still waiting for that lock, it checks whether it is in a deadlock situation
• If it is in a deadlock situation, it kills itself (is that true? or does it kill the other transaction? or some other selection criteria like the lower pid is killed or smthg?)
• If it is not in a deadlock situation, it will check after another deadlock_timeout seconds whether it is now in a deadlock situation

That is mostly correct. The timer for the deadlock check is set only once. If there is no deadlock after deadlock_timeout, there will be no more check for deadlocks. That is correct, because the only way that we could get into a deadlock later on is if another backend gets stuck behind a lock as well. Then that backend's deadlock check will detect and resolve the problem.

See the functions DeadLockCheck() and CheckDeadLock() in the source code.