Re: [PATCH] Fix NULL dereference in pg_get_database_ddl() - Mailing list pgsql-hackers

On Sat, 11 Apr 2026 at 01:58, Ayush Tiwari <ayushtiwari.slg01@gmail.com> wrote:
> Deterministic reproduction:
>
> CREATE DATABASE regression_testdb;
> SET allow_system_table_mods = on;
> UPDATE pg_database
>  SET dattablespace = 99999
>  WHERE datname = 'regression_testdb';
> RESET allow_system_table_mods;
>
> SELECT * FROM pg_get_database_ddl('regression_testdb');
>
> The attached patch fixes this by checking for NULL before calling
> pg_strcasecmp().  In that case, pg_get_database_ddl() simply omits the
> TABLESPACE clause.

Can you explain why this method of self-inflicted catalogue corruption
is any more important than any of the just-about-infinite other ways
there are of causing issues by manually updating the catalogue tables?

The typical response to this sort of thing can be seen in the thread
in [1], in particular, the response in [2].

David

[1] https://www.postgresql.org/message-id/19383-e6b60ec2a4fce5b0@postgresql.org
[2] https://www.postgresql.org/message-id/1538113.1768921841%40sss.pgh.pa.us

Thanks for the review.

I dug into this further and I don't think I can support the patch as a real bug fix.

My original thought was that there might be a supported concurrent-DDL path where
pg_get_database_ddl_internal() reads pg_database.dattablespace, and then a
concurrent change makes get_tablespace_name() return NULL before the second
lookup. I tried to reproduce that by widening the window and testing
ALTER DATABASE ... SET TABLESPACE together with DROP TABLESPACE, but I could not
make it happen.

The reason seems to be that both catalog reads are governed by the same
statement-level CatalogSnapshot, so the function continues to see a consistent
catalog view for the duration of the call. On the other side, DROP TABLESPACE
also doesn't appear able to commit a state where pg_database still points at a
vanished tablespace row through supported SQL, because the drop fails while the
database directory is still present.

So at this point the remaining crash case seems to require a broken catalog state,

which puts this in the same bucket as the precedent you cited rather than a
supported-path bug.

Given that, I don't think this patch is worth pursuing further in its current
form. I'll drop it here unless I find a supported reproducer or a different,
stronger issue around the missing database/tablespace bookkeeping.