On 2016-06-22 11:45:16 -0400, Tom Lane wrote:
> petrum@gmail.com writes:
> > File: postgresql-9.4.4/src/backend/replication/logical/reorderbuffer.c
> > Function: ReorderBufferInterTXNInit
> > Line: 870
>
> > The line is
> > if (txn->nentries != txn->nentries_mem)
> > But shouldn't be there cur_txn instead of txn?
>
> Actually, the function is ReorderBufferIterTXNInit, and in HEAD this
> is line 963, but yeah that looks pretty broken. Andres, do you
> concur?
Ugh, yes, that looks broken. In a way that can very likely lead to wrong
data being returned :(. I assume an empty toplevel transaction +
subtransactions with spilled-to-disk contents will be bad.
> Or maybe the logic needs to be different for subtransactions?
>
> > I do not know exactly the semantics of the code because I detected the
> > problem with a CodeSonar prototype plugin.
>
> Seems like a cool tool.
Indeed. What heuristic lead to detecting this? I can think of some, but
they all owuld have significant false-positive rates.