Re: pg_visibility's pg_check_visible() yields false positive when working in parallel with autovacuum

Hi,

On 2022-02-18 08:56:37 -0800, Andres Freund wrote:
> Could you try to minimize the script? A 300 line reproducer is quite long. And
> it looks like it won't even work in non postgres-pro tree.
>
> One thing to do would be to modify pg_visibility to elog(PANIC, "something")
> when it encounters corruption. Then you would have a chance of inspecting the
> state of the tuple/page in that moment.

Oh, I have been able to reliably reproduce this on HEAD. I modified
record_corrupt_item() to PANIC and then:

psql regression:
  BEGIN ;SELECT txid_current();
  <leave open>

psql postgres
  DROP TABLE IF EXISTS vacuum_test_0;
  create table vacuum_test_0 as select 42 i;
  vacuum (disable_page_skipping) vacuum_test_0;
  select * from pg_check_visible('vacuum_test_0');

At which point there immediately is a crash.


This reproduces in earlier versions too, at least back to 10.


I *think* this is a false positive:

- PGPROC->xmin is computed without regard for the database in which the other
  sessions are running. Due to the the txid_current() session this includes an
  older xid.

- During the VACUUM in vis.sql the only connection to the database pgbench
  connects to is VACUUM and thus ignored when determining horizons (due to
  PROC_IN_VACUUM). Therefore the horizon is computed to latestCompletedXid +
  1.

- But during pg_check_visible(), the current session is *not* marked as
  PROC_IN_VACUUM. So the horizon is the xid from the txid_current().

Boom.

Greetings,

Andres Freund