Hi,
On 2021-08-03 10:59:25 +1200, David Rowley wrote:
> On Sat, 31 Jul 2021 at 08:38, Andres Freund <andres@anarazel.de> wrote:
> > There is at least one path using tuplecontext that reaches code that
> > could end up freeing memory to a significant enough degree to care about
> > generation.c effectively not using that memory:
> > tuplesort_putdatum()->datumCopy()->EOH_flatten_into()
> > On a quick look I didn't find any expanded record user that frees
> > nontrivial amounts of memory, but I didn't look all that carefully.
>
> I guess we could just use a normal context for datum sorts if we
> thought that might be a problem.
I think that's probably a cure worse than the disease. I suspect datum sorts
can benefit from the higher density quite a bit...
> I'm not too familiar with the expanded object code, but I'm struggling
> to imagine why anything would need to do a pfree in there. We just do
> EOH_get_flat_size() to determine how big to make the allocation then
> allocate some memory for EOH_flatten_into() to use to expand the
> object into.
I can see some scenarios with a bit more creative uses of expanded
objects. We've e.g. been talking about using EA to avoid repeated and partial
detoasting overhead and you might need to do some more toast fetches when
flattening. Toast fetches always allocate, and if the fetch were only for
later parts of the tuple, the fetched data would need to be freed.
It's probably fine to deal with this at later time, and just leave a comment
somewhere.
It could be addressed by having a bump style allocator combined with having
freelists. It's not like the tuplesort.c case is actually interested in the
generational behaviour of generation.c (which makes freelists uninteresting),
it's just that generation.c is the densest allocator that we have right now...
Greetings,
Andres Freund
