Re: Failures with wal_consistency_checking and 13~

On Sat, Jun 20, 2020 at 1:16 PM Alexander Korotkov
<a(dot)korotkov(at)postgrespro(dot)ru> wrote:
> On Fri, Jun 19, 2020 at 10:34 PM Alvaro Herrera
> <alvherre(at)2ndquadrant(dot)com> wrote:
> >
> > On 2020-Jun-15, Michael Paquier wrote:
> >
> > > I have begun my annual study of WAL consistency across replays, and
> > > wal_consistency_checking = 'all' is pointing out at some issues with
> > > at least VACUUM and SPGist:
> > > FATAL: inconsistent page found, rel 1663/16385/22133, forknum 0,
> > > blkno 15
> > > CONTEXT: WAL redo at 0/739CEDE8 for SPGist/VACUUM_REDIRECT: newest
> > > XID 4619
> > >
> > > It may be possible that there are other failures, I have just run
> > > installcheck and this is the first failure I saw after replaying all
> > > the generated WAL on a standby. Please note that I have also checked
> > > 12, and installcheck passes.
> >
> > Umm. Alexander, do you an idea of what this is about?
>
> I don't have idea yet, but I'll check this out

I have discovered and fixed the issue in a44dd932ff. spg_mask()
masked unused space only when pagehdr->pd_lower >
SizeOfPageHeaderData. But during the vacuum regression tests, one
page has been erased completely and pagehdr->pd_lower was set to
SizeOfPageHeaderData. Actually, 13 didn't introduce any issue, it
just added a test that spotted the issue. The issue is here since
a507b86900.

------
Regards,
Alexander Korotkov