Date: Wed, 28 Feb 2018 20:30:19 -0300
From: Alvaro Herrera <alvherre@alvh.no-ip.org>
To: Alexandre Garcia <alexandre@vmfarms.com>
Cc: Andres Freund <andres@anarazel.de>,
	"pgsql-admin@postgresql.org" <pgsql-admin@postgresql.org>
Subject: Re: postgresql 9.6 - cannot freeze committed xmax
Message-ID: <20180228233019.qydod7ez3xbgn5ot@alvherre.pgsql>
MIME-Version: 1.0
Content-Type: multipart/mixed; boundary="saqj6tdvqhp5hsoj"
Content-Disposition: inline
Content-Transfer-Encoding: 8bit
In-Reply-To: 
 <CAPYLKR4BqAa-FBM6QoEj8OYZXafu9RUEusoecrpo0ZoBxoY4Wg@mail.gmail.com>
User-Agent: NeoMutt/20170306-137-4415bd-dirty (1.8.0)
Precedence: bulk


--saqj6tdvqhp5hsoj
Content-Type: text/plain; charset=iso-8859-1
Content-Disposition: inline
Content-Transfer-Encoding: 8bit

Alexandre Garcia wrote:
> >
> > Ooh.
> >
> > If you SELECT FOR UPDATE a tuple in 9.2, bit 0x0040 gets set in
> > infomask, and nothing else.  If you pg_upgrade and later try to freeze
> > such a tuple, it will fail with the error reported.
> >
> > The correct test to use is HEAP_XMAX_IS_LOCKED_ONLY, which also tests
> > for the above condition.
> >
> > I will verify this theory and push a patch shortly, if it proves
> > correct.
>
> Oh good news  :). I have my old 9.2 around if you need me to do more
> testing on it.

Not necessary -- I reproduced the problem (quite easily -- just SELECT
FOR UPDATE a tuple in 9.2, then pg_upgrade, then VACUUM FREEZE the table
in the upgraded server) and confirm that it doesn't happen in 9.6.6,
happens in 9.6.8 (didn't try 9.6.7 but code is the same as 9.6.8), and
is fixed with the attached patch.

Will push soon ... probably tomorrow as I have to leave the building
now.

Cheers

-- 
Álvaro Herrera                https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

--saqj6tdvqhp5hsoj
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment; filename="0001-Fix-freeze-xmax-bug.patch"