Horton

Thread

Commits

  1. Ensure we have a snapshot while dropping ON COMMIT DROP temp tables.

  1. Assert failure when CREATE TEMP TABLE

    Richard Guo <guofenglinux@gmail.com> — 2023-10-16T11:02:29Z

    I came across an Assert failure while running sqlancer.  The repro query
is

CREATE TEMP TABLE t0 (c0 int4range CHECK ((('')||(0.005014271491241007) NOT
IN (upper(''), ((' 9O')||FALSE))) BETWEEN SYMMETRIC
(initcap('dj')LIKE(((('EizW~')||t0.c0) COLLATE "en_AG.utf8"))) AND
((((0.18773675)::MONEY) IN (0.52246463::MONEY))AND(t0.c0 BETWEEN SYMMETRIC
((t0.c0+t0.c0)) AND t0.c0)))) USING heap ON COMMIT DROP;

The failure happens in init_toast_snapshot():

 /*
  * Catalog snapshots can be returned by GetOldestSnapshot() even if not
  * registered or active. That easily hides bugs around not having a
  * snapshot set up - most of the time there is a valid catalog snapshot.
  * So additionally insist that the current snapshot is registered or
  * active.
  */
 Assert(HaveRegisteredOrActiveSnapshot());

And 'git bisect' says that the first bad commit is 2489d76c, which makes
me confused because the problemed query does not seem to involve
nullingrels stuff.

Thanks
Richard

  2. Re: Assert failure when CREATE TEMP TABLE

    Tom Lane <tgl@sss.pgh.pa.us> — 2023-10-16T15:46:54Z

    Richard Guo <guofenglinux@gmail.com> writes:
> I came across an Assert failure while running sqlancer.  The repro query
> is

> CREATE TEMP TABLE t0 (c0 int4range CHECK ((('')||(0.005014271491241007) NOT
> IN (upper(''), ((' 9O')||FALSE))) BETWEEN SYMMETRIC
> (initcap('dj')LIKE(((('EizW~')||t0.c0) COLLATE "en_AG.utf8"))) AND
> ((((0.18773675)::MONEY) IN (0.52246463::MONEY))AND(t0.c0 BETWEEN SYMMETRIC
> ((t0.c0+t0.c0)) AND t0.c0)))) USING heap ON COMMIT DROP;

Hmm.  AFAICT, the critical thing about this CHECK condition is simply
that it's bulky enough to create an out-of-line toasted "conbin" field
in pg_constraint.  When we come to ON COMMIT DROP, we have to delete
the pg_constraint entry and thereby the toast item, and we evidently
lack a snapshot at that point.

> And 'git bisect' says that the first bad commit is 2489d76c, which makes
> me confused because the problemed query does not seem to involve
> nullingrels stuff.

Probably the reason is that addition of nullingrels fields to the Vars
pushed it over the edge of needing to be out-of-line.  The test case
is very close to the line as stated --- for example, I found that it
didn't crash if I changed "en_AG.utf8" to "C", and wasted some time
pursuing the idea that the collation had something to do with it.
The crash got much more stable after adding a couple more dummy clauses
to the CHECK condition, and I've also reproduced it with clauses as
straightforward as "c0 < 'very-long-constant'".

			regards, tom lane

  3. Re: Assert failure when CREATE TEMP TABLE

    Alexander Law <exclusion@gmail.com> — 2023-10-17T09:00:00Z

    16.10.2023 18:46, Tom Lane wrote:
> Richard Guo <guofenglinux@gmail.com> writes:
>> I came across an Assert failure while running sqlancer.  The repro query
>> is
> Hmm.  AFAICT, the critical thing about this CHECK condition is simply
> that it's bulky enough to create an out-of-line toasted "conbin" field
> in pg_constraint.  When we come to ON COMMIT DROP, we have to delete
> the pg_constraint entry and thereby the toast item, and we evidently
> lack a snapshot at that point.

That's very similar to bug #18127, and looks like one more case to the
collection [1], with RemoveConstraintById() -> CatalogTupleDelete() ->
simple_heap_delete() ...

https://www.postgresql.org/message-id/9541c507-a325-7768-38b0-90f6532c01e3%40gmail.com

Best regards,
Alexander

  4. Re: Assert failure when CREATE TEMP TABLE

    Richard Guo <guofenglinux@gmail.com> — 2023-10-18T06:57:37Z

    On Mon, Oct 16, 2023 at 11:46 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:

> Richard Guo <guofenglinux@gmail.com> writes:
> > And 'git bisect' says that the first bad commit is 2489d76c, which makes
> > me confused because the problemed query does not seem to involve
> > nullingrels stuff.
>
> Probably the reason is that addition of nullingrels fields to the Vars
> pushed it over the edge of needing to be out-of-line.  The test case
> is very close to the line as stated --- for example, I found that it
> didn't crash if I changed "en_AG.utf8" to "C", and wasted some time
> pursuing the idea that the collation had something to do with it.
> The crash got much more stable after adding a couple more dummy clauses
> to the CHECK condition, and I've also reproduced it with clauses as
> straightforward as "c0 < 'very-long-constant'".


Ah, this clears up my confusion.  Thanks for the fix.

Thanks
Richard