Re: pgsql: Improve accounting for memory used by shared hash tables

On 4/4/25 02:41, Tomas Vondra wrote:
> On 4/4/25 01:43, Tomas Vondra wrote:
>> On 4/4/25 00:57, David Rowley wrote:
>>> On Thu, 3 Apr 2025 at 04:16, Tomas Vondra <tomas(dot)vondra(at)postgresql(dot)org> wrote:
>>>> Improve accounting for memory used by shared hash tables
>>>
>>> I've not looked into why, but this is causing an issue in the
>>> join_rel_hash during add_join_rel(). See the attached script.
>>>
>>> ERROR: pfree called with invalid pointer 0x60a15edc44e0 (header
>>> 0x0000002000000008)
>>>
>>
>> Thanks for the report and reproducer. I'll take a look tomorrow.
>>
>
> I took a quick look, and I think the reason is fairly simple - the
> commit allocates the header and the directory as a single chunk. And for
> shared hash tables that's fine, because those have non-expandable
> directory. But the patch does the same thing for non-shared hash tables
> (not intentionally), which means that if we end up expanding the hash,
> it fails in dir_realloc(). Because hashp->dir is not a separately
> allocated chunk.
>
> This is clearly a bug in the patch, I should have caught this during a
> review. But I'm also quite surprised none of the regression tests seems
> to expand the hash table ...
>
> I'll think about a way to fix this tomorrow.
>

I ended up reverting this. Unfortunately, the patch assumed the
directory is pre-allocated and not expanding in more places. I wasn't
sure how long would it take me to fix this, or how invasive the fix
would be. It seems more appropriate to revert and then maybe apply a
reworked patch (not going to happen for PG18).

Thanks for the report, sorry for missing the issue in the first place.

--
Tomas Vondra