Date: Fri, 22 Nov 2019 00:47:01 +0100
From: Tomas Vondra <tomas.vondra@2ndquadrant.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: =?utf-8?Q?Ond=C5=99ej?= Jirman <ienieghapheoghaiwida@xff.cz>,
	pgsql-bugs@lists.postgresql.org
Subject: Re: BUG #16129: Segfault in tts_virtual_materialize in logical
 replication worker
Message-ID: <20191121234701.2wtyc77zd5ux3myf@development>
References: <20191121150716.szxqukxrtpjpckm3@core.my.home>
 <20191121152732.hu6y4wzw7nkxpsvk@development>
 <20191121155707.pqkgjphh65o5fapz@core.my.home>
 <20191121161502.xfiy65i3cm6adscr@core.my.home>
 <20191121165752.dffge6bh756xlfdg@development>
 <20191121173555.midj6vp4dmiiqs24@core.my.home>
 <20191121185844.6qsqhwzzuhu3fpsg@core.my.home>
 <20191121214115.3tgefmvkmrzaim74@development>
 <20191121220854.ebdzjtloel56sfwu@development>
 <13122.1574375845@sss.pgh.pa.us>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii; format=flowed
Content-Disposition: inline
In-Reply-To: <13122.1574375845@sss.pgh.pa.us>
Precedence: bulk

On Thu, Nov 21, 2019 at 05:37:25PM -0500, Tom Lane wrote:
>Tomas Vondra <tomas.vondra@2ndquadrant.com> writes:
>> FWIW my hunch is the bug is somewhere in this chunk of code from
>> apply_heap_update:
>
>>          oldctx = MemoryContextSwitchTo(GetPerTupleMemoryContext(estate));
>>          ExecCopySlot(remoteslot, localslot);
>>          slot_modify_cstrings(remoteslot, rel, newtup.values, newtup.changed);
>>          MemoryContextSwitchTo(oldctx);
>
>My first reaction to looking at this was that ExecCopySlot() could
>be dropped entirely, because the first thing that slot_modify_cstrings
>does is
>
>	slot_getallattrs(slot);
>	ExecClearTuple(slot);
>
>That slot_getallattrs call seems 100% useless as well, because
>surely ExecClearTuple is just going to drop all the data.
>

I don't think that's quite true. After the ExecCopySlot call, the
pass-by-ref Datums in remoteslot will point to a tuple attached to
localslot. But it does not pass the tuple 'ownership' to the remoteslot,
i.e. the flag TTS_FLAG_SHOULDFREE won't be set, i.e. the tuple won't be
freed.

>On closer inspection, though, it looks like the author of this
>code is relying on ExecClearTuple to *not* destroy the data,
>because the subsequent loop might overwrite only some of the
>slot's columns, before it does
>
>	ExecStoreVirtualTuple(slot);
>
>which supposes that all the columns are valid.
>
>But of course this is 100% broken, because whatever ExecClearTuple
>may have done or not done to the slot's datum/isnull arrays, it
>certainly pfree'd the slot's physical tuple.  So any pass-by-ref
>datums are now dangling pointers.
>
>I imagine the only reason this code has gotten past the valgrind
>animals is that we're not testing cases where non-replaced columns
>in the subscriber table are of pass-by-ref types.
>

I haven't checked, but I'd imagine we actually do such tests. I've
however tried to reproduce this, unsuccessfully.

>I draw a direct line between the existence of this bug and the lack
>of commentary in slot_modify_cstrings about what it's doing.  If the
>author had troubled to write a comment explaining these machinations,
>he might've noticed the bug, or at least reviewers might have.
>

Not sure. More comments would not hurt, but I don't see any bug yet.


regards

-- 
Tomas Vondra                  http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services