On Mon, Feb 21, 2011 at 6:31 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> I wrote:
>> Ugh. =A0That quick little "ExecRemoveJunk" is a lot more dangerous than =
it
>> looks. =A0I had actually looked at this before, but concluded it was OK
>> because I couldn't reproduce the problem with a trigger in place.
>> I guess I wasn't unlucky enough to have the chance pointer equality
>> occur.
>
> On closer inspection, the palloc collision is actually extremely likely,
> because what will happen is we'll pfree the old tuple and immediately
> palloc the new one, and if the new one is of sufficiently similar size
> that it needs the same size of alloc chunk, it's *guaranteed* to get
> that same chunk back, because of the LIFO free-chunk chains in aset.c.
>
> The reason that the problem is hard to reproduce is that most triggers
> (certainly those written in plpgsql) will give back newly allocated
> tuples even when you return the unmodified NEW tuple. =A0The only way to
> expose the problem is for ExecBRUpdateTrigger's trigger-calling loop to
> not replace the "newtuple", and the easiest way for that to happen is if
> all the triggers are disabled. =A0So that's why you're seeing it when
> fooling with the replication-role setting. =A0I was able to reproduce the
> failure by creating a trigger with a false WHEN condition, and of course
> there are other ways to prevent a trigger from being called too.
I bumped into the 'virtual tuple' error on 9.0.3 on a completely
unrelated query:
create table foo(a int, b int);
insert into foo select 1, 1 where not exists (select 1 from foo where
a =3D 1 for update);
9.0.4 got rid of the error, but as there are no triggers or anything
like that on the foo so I thought I'd post my findings here in case
anyone else bumped into it.
merlin
