Re: PostgreSQL12 crash bug report

Andres Freund <andres(at)anarazel(dot)de> writes:
> On 2019-07-31 19:07:29 -0400, Tom Lane wrote:
>> It's setting up an es_epqTupleSlot[] entry on the assumption that it
>> should have the same tupdesc as the output tuple that's to be rechecked.
>> This might accidentally fail to malfunction in single-table cases,
>> but it's quite broken for any join case --- in particular, for the
>> given test case, the scan tuple for the VALUES node surely doesn't have
>> the same tupdesc as the join result.

> To make sure I understand - the problem isn't the slot that we've
> created in nodeModifyTable.c via EvalPlanQualSlot(), right? It's the one
> we create in EvalPlanQualFetchRowMarks(), because we don't have a proper
> tuple type handy to create the slot with?

Yeah, I think nodeModifyTable.c is fine, because it always passes the
target relation. EvalPlanQualFetchRowMarks is not fine, and I'm unsure
about the call in nodeLockRows.c.

> Previously we simply didn't need to know the type during EPQ setup,
> because we only stored a HeapTuple anyway. And we'd know the appropriate
> tupledesc at the places that access the tuple.

Right. So we gotta refactor that somehow.

> One bigger change - but possibly one worth it - would be to basically
> move the work done in EvalPlanQualFetchRowMarks() to be done on-demand,
> at least for ROW_MARK_COPY.

Hm. Too tired to think that through right now.

regards, tom lane