Corner-case bug in pg_rewind

Hi

Take the following cluster with:
  - node1 (initial primary)
  - node2 (standby)
  - node3 (standby)

Following activity takes place (greatly simplified from a real-world situation):

1. node1 is shut down.
2. node3 is promoted
3. node2 is attached to node3.
4. node1 is attached to node3
5. node1 is then promoted (creating a split brain situation with
   node1 and node3 as primaries)
6. node2 and node3 are shut down (in that order).
7. pg_rewind is executed to reset node2 so it can reattach
   to node1 as a standby. pg_rewind claims:

    pg_rewind: servers diverged at WAL location X/XXXXXXX on timeline 2
    pg_rewind: no rewind required

8. based off that assurance, node2 is restarted with replication configuration
   pointing to node1 - but it is unable to attach, with node2's log reporting
   something like:

      new timeline 3 forked off current database system timeline 2
before current recovery point X/XXXXXXX

The cause is that pg_rewind is assuming that if the node's last
checkpoint matches the
divergence point, no rewind is needed:

    if (chkptendrec == divergerec)
        rewind_needed = false;

but in this case there *are* records beyond the last checkpoint, which can be
inferred from "minRecoveryPoint" - but this is not checked.

Attached patch addresses this. It includes a test, which doesn't make use of
the RewindTest module, as that hard-codes a primary and a standby, while here
three nodes are needed (I can't come up with a situation where this can be
reproduced with only two nodes). The test sets "wal_keep_size" so would need
modification for Pg12 and earlier.

Regards

Ian Barwick

-- 
  Ian Barwick                   https://www.2ndQuadrant.com/
  PostgreSQL Development, 24x7 Support, Training & Services