Re: memory leak in trigger handling (since PG12)

On 5/24/23 17:37, Tom Lane wrote:
> Tomas Vondra <tomas(dot)vondra(at)enterprisedb(dot)com> writes:
>> While looking for other places allocating stuff in ExecutorState (for
>> the UPDATE case) and leaving it there, I found two more cases:
>
>> 1) copy_plpgsql_datums
>
>> 2) make_expanded_record_from_tupdesc
>> make_expanded_record_from_exprecord
>
>> All of this is calls from plpgsql_exec_trigger.
>
> Can you show a test case in which this happens? I added some
> instrumentation and verified at least within our regression tests,
> copy_plpgsql_datums' CurrentMemoryContext is always plpgsql's
> "SPI Proc" context, so I do not see how there can be a query-lifespan
> leak there, nor how your 0003 would fix it if there is.
>

Interesting. I tried to reproduce it, but without success, and it passes
even with an assert on the context name. The only explanation I have is
that the gdb script I used might have been a bit broken - it used
conditional breakpoints like this one:

break AllocSetAlloc if strcmp(((MemoryContext) $rdi)->name, \
"ExecutorState") == 0
commands
bt
cont
end

but I just noticed gdb sometimes complains about this:

Error in testing breakpoint condition:
'__strcmp_avx2' has unknown return type; cast the call to its declared
return type

The breakpoint still fires all the commands, which is pretty surprising
behavior, but that might explain why I saw copy_plpgsql_data as another
culprit. And I suspect the make_expanded_record calls might be caused by
the same thing.

I'll check deeper tomorrow, when I get access to the original script
etc. We can ignore these cases until then.

Sorry for the confusion :-/

regards

--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company