Tom Lane wrote:
>
> "Hiroshi Inoue" <Inoue@tpf.co.jp> writes:
> > It seems that init_irels() should be called after
> > InitializeTransactionSystem() was called.
>
> Can we just swap the order of the RelationCacheInitialize() and
> InitializeTransactionSystem() calls in InitPostgres? If that
> works, I'd have no objection.
>
It doesn't work. InitializeTransactionSystem() requires
pg_log/pg_variable relations which are already built in
RelationCacheInitialize(). A few critical relations
including pg_log/pg_variable are built in RelationCache
Initialize() without touching database. It's OK but
init_irels() touches system tables to build a few
critical index relations. IMHO init_irels() should
be separated from RelationCacheInitialize().
In the meantime,I have another anxiety. init_irels()
(RelationCacheInitialize()) seems to be called while
Locking is disabled. This seems to mean that init_irels()
could access to system tables even when they are in
vacuum. HeapTupleSatisfiesXXXX() doesn't seem to take
such cases into account except HeapTupleSatisfiesDirty().
HeapTupleSatisfiesXXXX() sets HEAP_XMIN_COMMITTED or
HEAP_XMIN_INVALID mask for HEAP_MOVED_IN(OFF) tuples.
Regards.
Hiroshi Inoue