Re: Fix a server crash problem from pg_get_database_ddl

Hi,

Adding Tom to the thread explicitly to seek his opinion.

On Wed, Apr 15, 2026 at 6:36 PM Chao Li <li.evan.chao@gmail.com> wrote:

>
>
> > On Apr 16, 2026, at 09:23, Japin Li <japinli@hotmail.com> wrote:
> >
> > On Wed, 15 Apr 2026 at 20:44, "Jack Bonatakis" <jack@bonatak.is> wrote:
> >> I have reproduced this error against the current master:
> >>
> >> ```
> >> CREATE TABLESPACE ts1 LOCATION '/workspace/tablespaces/pg_bug_ts1';
> >> CREATE DATABASE db1 TABLESPACE ts1;
> >> DELETE FROM pg_tablespace WHERE spcname = 'ts1';
> >> SELECT * FROM pg_get_database_ddl('db1'::regdatabase);
> >>
> >> server closed the connection unexpectedly
> >> This probably means the server terminated abnormally
> >> before or while processing the request.
> >> The connection to the server was lost. Attempting reset: Failed.
> >> ```
> >> Backend logs show:
> >>
> >> ```
> >> [1] LOG:  client backend (PID 15420) was terminated by signal 11:
> Segmentation fault
> >> [1] DETAIL:  Failed process was running: SELECT * FROM
> pg_get_database_ddl('db1'::regdatabase);
> >> [1] LOG:  terminating any other active server processes
> >> ```
> >> After applying the patch:
> >>
> >> ```
> >> SELECT * FROM pg_get_database_ddl('db1'::regdatabase);
> >> ERROR:  tablespace with OID 16393 does not exist
> >> HINT:  To recover, try ALTER DATABASE ... SET TABLESPACE ... to a valid
> tablespace.
> >> ```
> >> and backend logs show:
> >>
> >> ```
> >> [56] ERROR:  tablespace with OID 16393 does not exist
> >> [56] HINT:  To recover, try ALTER DATABASE ... SET TABLESPACE ... to a
> valid tablespace.
> >> [56] STATEMENT:  SELECT * FROM pg_get_database_ddl('db1'::regdatabase);
> >> ```
> >> All tests pass.
> >>
> >> The only note I'd have on the code change is that there is no
> accompanying test. It seems like a TAP test would be
> >> reasonable, but I am quite new and will defer to whether you think
> that's the right call or even necessary.
> >>
> >> Jack
> >
> > This seems similar to [1]. Could you please confirm?
> >
> > [1]
> https://www.postgresql.org/message-id/CAJTYsWXcd324VELk%3D9KdsfTsua9So3Yexqv7N3B23h9zAUD40g%40mail.gmail.com
> .
> >
> > --
> > Regards,
> > Japin Li
> > ChengDu WenWu Information Technology Co., Ltd.
> >
> >
>
> Thanks for printing out that. Yes, they are similar.
>
> I agree with what Tom said in [2]:
> ```
> This is not a bug. This is a superuser intentionally breaking
> the system by corrupting the catalogs. There are any number
> of ways to cause trouble with ill-advised manual updates to a
> catalog table. Try, eg, "DELETE FROM pg_proc" (... but not in
> a database you care about).
> ```
>
> So, let me take back this patch.
>
> [2] https://www.postgresql.org/message-id/1538113.1768921841@sss.pgh.pa.us


In this case, it is a very corner case but not something superuser
intentionally breaks.
For example, a concurrent tablespace drop + database ddl to assign a
different tablespace or default.
We aren't acquiring Access Share lock on the DB in this function
(intentional) so it is a good practice
to do the null checks. Of course, it makes more sense to add this comment
while doing a code review.
I will let Tom and others chime in with their thoughts on fixing this.

Attached an injection point test to show the race. Not intended to commit.

Thanks,
Satya