On Tue, Nov 28, 2023 at 07:00:16PM +0100, David Geier wrote:
> PostgreSQL hit the following assertion during error cleanup, after being OOM
> in dsa_allocate0():
>
> void dshash_detach(dshash_table *hash_table) {
> ASSERT_NO_PARTITION_LOCKS_HELD_BY_ME(hash_table);
>
> called from pgstat_shutdown_hook(), called from shmem_exit(), called from
> proc_exit(), called from the exception handler.
Nice find.
> AutoVacWorkerMain() pgstat_report_autovac() pgstat_get_entry_ref_locked()
> pgstat_get_entry_ref() dshash_find_or_insert() resize() resize() locks all
> partitions so the hash table can safely be resized. Then it calls
> dsa_allocate0(). If dsa_allocate0() fails to allocate, it errors out. The
> exception handler calls proc_exit() which normally calls LWLockReleaseAll()
> via AbortTransaction() but only if there's an active transaction. However,
> pgstat_report_autovac() runs before a transaction got started and hence
> LWLockReleaseAll() doesn't run before pgstat_shutdown_hook() is called.
From a glance, it looks to me like the problem is that pgstat_shutdown_hook
is registered as a before_shmem_exit callback, while ProcKill is registered
as an on_shmem_exit callback. However, IIUC even moving them to the same
list wouldn't be sufficient because the pg_stat_shutdown_hook is registered
after ProcKill, and the code that calls the callbacks walks backwards
through the list.
I would expect your patch to fix this particular issue, but I'm wondering
whether there's a bigger problem here.
--
Nathan Bossart
Amazon Web Services: https://aws.amazon.com
