From: Tom Lane <tgl@sss.pgh.pa.us>
To: Andres Freund <andres@anarazel.de>
cc: Michael Paquier <michael@paquier.xyz>,
        Justin Pryzby <pryzby@telsasoft.com>,
        PostgreSQL Hackers <pgsql-hackers@postgresql.org>
Subject: Re: ENOSPC FailedAssertion("!(RefCountErrors == 0)"
In-reply-to: <20180716131337.usestwjviwwocpyz@alap3.anarazel.de>
References: <20180627233939.GA10276@telsasoft.com>
 <20180628020432.GB11054@paquier.xyz> <7317.1531694923@sss.pgh.pa.us>
 <20180716131337.usestwjviwwocpyz@alap3.anarazel.de>
Comments: In-reply-to Andres Freund <andres@anarazel.de>
	message dated "Mon, 16 Jul 2018 06:13:37 -0700"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <20435.1531751966.1@sss.pgh.pa.us>
Date: Mon, 16 Jul 2018 10:39:26 -0400
Message-ID: <20436.1531751966@sss.pgh.pa.us>
Precedence: bulk

Andres Freund <andres@anarazel.de> writes:
> On 2018-07-15 18:48:43 -0400, Tom Lane wrote:
>> So basically, WAL replay hits an error while holding a buffer pin, and
>> nothing is done to release the buffer pin, but AtProcExit_Buffers thinks
>> something should have been done.

> I think there's a few other cases where we hit this. I've seen something
> similar from inside checkpointer / BufferSync(). I'd be surprised if
> bgwriter couldn't be triggered into the same.

Hm, yeah, on reflection it's pretty obvious that those are hazard cases.

> I'm pretty sure that we do *not* force a panic on all nonzero-exit-code
> cases for other subprocesses.

That's my recollection as well -- mostly, we just start a new one.

So I said I didn't want to do extra work on this, but I am looking into
fixing it by having these aux process types run a ResourceOwner that can
be told to clean up any open buffer pins at exit.  We could be sure the
coverage is complete by dint of removing the special-case code in
resowner.c that allows buffer pins to be taken with no active resowner.
Then CheckForBufferLeaks can be left as-is, ie something we do only
in assert builds.

			regards, tom lane