Alvaro Herrera <alvhe...@alvh.no-ip.org> writes:
> ... This is because ColumnDef->inhcount is a 32-bit int, but
> Form_pg_attribute->attinhcount is int16, so we didn't break the overflow
> test for ColumnDef inhcount, but attinhcount has overflowed during
> assignment.
Ugh ... somebody's ancient oversight there. Or maybe attinhcount
was once int32, and we narrowed it for space reasons?
> From branch master, I propose we change those two members to int16
> (ColumnDef->inhcount and CookedConstraint->inhcount) to make those
> counters consistently use the same type; and then use
> pg_add_s16_overflow() instead of ++ for the increments, as in the
> attached patch. With this patch, the child table creation fails as
> expected ("too many inheritance parents").
+1. I didn't check if there were any other places to touch, but
this looks like a good solution for master.
> In stable branches, I see two possible approaches: we could use the same
> ptach as master (but add another int16 to the struct as padding, to
> avoid changing the struct layout),
That would not preserve ABI on machines with the wrong endianness.
> or, less intrusive, we could leave
> that alone and instead change the "overflow" after the addition to test
> inhcount > PG_INT16_MAX instead of < 0. Or we could leave it all alone.
On the whole I'd leave it alone in back branches. Nobody who's not
intentionally trying to break their table will hit this.
> (I'm not terribly enthused about adding a test that creates a child
> table with 2^16 parents, because of the added runtime -- on my machine
> the scripts above take about 4 seconds.)
Agreed, too expensive for the value.
regards, tom lane
