I wrote:
> * We now recursively enter ScanPgRelation, which (again) needs to do a
> search using pg_class_oid_index, so it (again) opens and locks that.
> BUT: LockRelationOid sees that *this process already has share lock on
> pg_class_oid_index*, so it figures it can skip AcceptInvalidationMessages.
BTW, I now have a theory for why we suddenly started seeing this problem
in mid-June: commits a54e1f158 et al added a ScanPgRelation call where
there had been none before (in RelationReloadNailed, for non-index rels).
That didn't create the problem, but it probably increased the odds of
seeing it happen.
Also ... isn't the last "relation->rd_isvalid = true" in
RelationReloadNailed wrong? If it got cleared during ScanPgRelation,
I do not think we want to believe that we got an up-to-date row.
regards, tom lane