Re: Fix race in ReplicationSlotRelease for ephemeral slots

On Thu, Jun 18, 2026 at 11:20 AM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
>
> On Wed, Jun 17, 2026 at 12:59 PM Xuneng Zhou <xunengzhou(at)gmail(dot)com> wrote:
> >
> > On Tue, Jun 16, 2026 at 8:46 PM Fujii Masao <masao(dot)fujii(at)gmail(dot)com> wrote:
> > >
> > > On Fri, Jun 12, 2026 at 7:54 PM Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> > > > I feel even if there is an argument to do such a refactoring, it can
> > > > be done separately. We can push forward with 0001 and then do more
> > > > discussion for 0002, if required. I can take care of 0001 unless
> > > > Fujii-San wishes to take care of it?
> > >
> > > Yeah, please feel free to work on 0001.
> > >
> > > Regarding 0002, since the race is very rare and non-fatal, I'm okay
> > > with accepting the risk rather than adding more refactoring just to
> > > avoid it.
> > >
> > > I'm a bit tempted to add a source comment explaining the risk and
> > > why we accept it, though, so other developers can understand
> > > the tradeoff. For example:
> > >
> > > diff --git a/src/backend/replication/logical/slotsync.c
> > > b/src/backend/replication/logical/slotsync.c
> > > index 05637344363..ca49f20e7d9 100644
> > > --- a/src/backend/replication/logical/slotsync.c
> > > +++ b/src/backend/replication/logical/slotsync.c
> > > @@ -560,6 +560,12 @@ drop_local_obsolete_slots(List *remote_slot_list)
> > > * the same shared memory as that of
> > > 'local_slot'. Thus check if
> > > * local_slot is still the synced one before
> > > performing the actual
> > > * drop.
> > > + *
> > > + * Because local_slot still points to a
> > > reusable slot-array entry,
> > > + * fields such as name or database OID could
> > > already be stale here.
> > > + * That could cause an incorrect cleanup
> > > decision for this cycle or
> > > + * briefly lock an unrelated database. We
> > > accept that risk because
> > > + * this race is rare and non-fatal.
> > > */
> > > SpinLockAcquire(&local_slot->mutex);
> > > synced_slot = local_slot->in_use &&
> > > local_slot->data.synced;
> >
> > Thanks for suggesting the comment! It helps to clarify the situation
> > and the trade-off we made here. I tweaked it a bit and added it to the
> > patches prepared by Zhijie.
> >
>
> + *
> + * We cannot close this window by holding
> + * ReplicationSlotControlLock while taking the database lock,
> + * because the database-drop path holds the database lock and then
> + * scans replication slots.
>
> The database-drop path acquires ReplicationSlotControlLock in shared
> mode, so not sure if the above is completely correct, here you are
> going in the direction of trying to defend that no easy solution
> exists which needs more thought.

I agree that this phrasing could be misleading. The dangerous part is
in the later actual slot drop. Eventually the database-drop path needs
ReplicationSlotControlLock in exclusive mode to set slot->in_use =
false. If the slotsync holds ReplicationSlotControlLock in shared mode
while it tries to take DB AccessShareLock. Then this would lead to a
deadlock. On second thought, I’m inclined not to add a comment
explaining the constraint here, since it would be difficult to do so
concisely and it's not directly related to the issue.

> Fujii-San's proposal was better but
> there also we may need to be a bit more specific about "That could
> cause an incorrect cleanup decision ...", otherwise, it makes the
> comment unclear.

OK, how about elaborate it a bit like this:

/*
* In the small window between getting the slot to drop and
* locking the database, there is a possibility of a parallel
* database drop by the startup process and the creation of a new
* slot by the user. This new user-created slot may end up using
* the same shared memory as that of 'local_slot'.
*
* If that happens, local_slot now describes the replacement slot:
* local_sync_slot_required() may have made its drop decision using
* the replacement slot's name or invalidation state, and slot_database
* may refer to the replacement slot's database. Thus check if
* local_slot is still a synced slot before performing the actual drop.
* This does not prove it is the original slot, but it prevents dropping
* an ordinary user-created replacement slot, and the copied database OID
* keeps lock/unlock symmetric. The remaining risk is limited to this
* cleanup cycle, such as briefly holding an unrelated database lock, and
* is acceptable here because this race is rare.
*/

> I am planning to commit and backpatch the fix for the first problem
> based on what Hou-San has shared (v2-*), then we can discuss how to
> improve the existing comments and if we agree on something, that can
> be a HEAD-only patch.

Thanks!

--
Regards,
Xuneng Zhou
HighGo Software Co., Ltd.