Hi Henson,

> Hi Tatsuo,
> 
> == 1. DEFINE evaluation use-after-free (to Tatsuo) ==
> 
>>> nocfbot-0039 (the DEFINE memory-leak fix) added a ResetExprContext() in
>>> update_reduced_frame, but it resets the wrong context.
>>>
>>> ps_ExprContext is the per-output-tuple context that ExecWindowAgg resets
>>> once per output row.  update_reduced_frame now resets it once per NFA
> row,
>>> while the output row is still being formed -- so a pass-by-ref window
>>> function result already datum-copied into that per-tuple memory (when
>>> numfuncs > 1) is freed before ExecProject reads it.
>>>
>>> Neither v47 nor the patch is the answer on its own: v47 had no reset
> here,
>>> so no use-after-free, but the DEFINE scratch accumulated over the whole
>>> forward scan (the leak nocfbot-0039 fixed); nocfbot-0039 added the
> per-row
>>> reset but on the shared per-output-tuple context.  We do want a per-row
>>> reset -- just not on that context.
>>>
>>> So I think this needs a dedicated ExprContext for DEFINE evaluation,
> reset
>>> once per NFA row: it keeps the memory bounded without touching the
>>> per-output-tuple results.
>>>
>>> Question: does a dedicated DEFINE ExprContext look right to you?
>>
>> Can we use winstate->tmpcontext instead?
> 
> I don't think we can.  tmpcontext is idle where update_reduced_frame
> would reset it, but not *during* DEFINE evaluation: a DEFINE with NEXT()
> re-enters the spooler from inside its ExecEvalExpr (ExecEvalRPRNavSet
> -> window_gettupleslot -> spool_tuples), and spool_tuples resets
> winstate->tmpcontext (via ExecQualAndReset on partEqfunction) for every
> input row it pulls under a PARTITION BY.  ExecEvalRPRNavRestore parks a
> pass-by-ref nav result in that per-tuple memory to survive until the
> next reset.  To give one example,
> 
>   -- the cast forces a pass-by-ref result; a by-value bigint is safe
>   DEFINE b AS PREV(price::numeric) < NEXT(price::numeric)
> 
> frees PREV's result when NEXT spools the next row, before numeric_lt()
> reads it -- the same use-after-free as nocfbot-0039, just on a different
> context.  It is the normal forward-scan path (the NFA runs ahead of the
> spool), and after a tuplestore spill even NEXT-free DEFINEs hit it.
> 
> More generally, tmpcontext fits its existing users because each one sets
> its slots, evaluates, and resets within a shallow, straight-line region;
> it is best kept to that limited, low-depth usage.  DEFINE evaluation
> re-enters the spooler mid-expression, so it cannot honor that contract.
> 
> So I think the dedicated DEFINE ExprContext is the right call: a
> once-per-NFA-row reset is a third cadence, distinct from tmpcontext
> (per-input-tuple) and ps_ExprContext (per-output-tuple) -- the same
> one-context-per-cadence pattern nodeWindowAgg and nodeAgg already use.

Thanks for the explanation. Got it. We need the third ExprContext.

Regards,
--
Tatsuo Ishii
SRA OSS K.K.
English: http://www.sraoss.co.jp/index_en/
Japanese:http://www.sraoss.co.jp


Reply via email to