MIME-Version: 1.0
References: 
 <TY4PR01MB177184FF9EE916F577E1F554194082@TY4PR01MB17718.jpnprd01.prod.outlook.com>
 <CAFC+b6o-hD5VxVLZQovmHSYykF8Qzq3eiuBU-U1F_yR9-y6P_w@mail.gmail.com>
 <TY4PR01MB177180A7CE60BCDF286B1C6F594172@TY4PR01MB17718.jpnprd01.prod.outlook.com>
 <CABPTF7VyH1-W2xnDspECDEzFGQj=WTFpZBCqKfM11OAZa6gQHQ@mail.gmail.com>
 <CAHGQGwE+2WSqiAYgNJRkf_twdB+uRGozjjGhUn76vUKZ8dzbSA@mail.gmail.com>
 <CABPTF7VeA8szPv7LYDVY9_7LftV-HM8NFVQR2natPKmr73JW+A@mail.gmail.com>
 <TY4PR01MB1771887D33612C5A45F7E9CDF941E2@TY4PR01MB17718.jpnprd01.prod.outlook.com>
In-Reply-To: 
 <TY4PR01MB1771887D33612C5A45F7E9CDF941E2@TY4PR01MB17718.jpnprd01.prod.outlook.com>
From: Amit Kapila <amit.kapila16@gmail.com>
Date: Thu, 11 Jun 2026 14:52:24 +0530
Message-ID: 
 <CAA4eK1LqFBKCkX2eoX3iQPxJJnzWTaCpdh9zNotxuoG8BgjdtA@mail.gmail.com>
Subject: Re: Fix race in ReplicationSlotRelease for ephemeral slots
To: "Zhijie Hou (Fujitsu)" <houzj.fnst@fujitsu.com>
Cc: Xuneng Zhou <xunengzhou@gmail.com>, Fujii Masao <masao.fujii@gmail.com>,
	Srinath Reddy Sadipiralla <srinath2133@gmail.com>,
	PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Archived-At: 
 <https://www.postgresql.org/message-id/CAA4eK1LqFBKCkX2eoX3iQPxJJnzWTaCpdh9zNotxuoG8BgjdtA%40mail.gmail.com>
Precedence: bulk

On Sat, Jun 6, 2026 at 3:05=E2=80=AFPM Zhijie Hou (Fujitsu)
<houzj.fnst@fujitsu.com> wrote:
>
> On Friday, June 5, 2026 8:45 PM Xuneng Zhou <xunengzhou@gmail.com> wrote:
> > On Wed, Jun 3, 2026 at 8:03=E2=80=AFPM Fujii Masao <masao.fujii@gmail.c=
om> wrote:
> > >
> > > On Tue, Jun 2, 2026 at 3:00=E2=80=AFPM Xuneng Zhou <xunengzhou@gmail.=
com>
> > wrote:
> > >
> > >   /* Drop the local slot if it is not required to be retained. */
> > >   if (!local_sync_slot_required(local_slot, remote_slot_list))
> > >   {
> > > + bool dropped =3D false;
> > > + NameData slot_name =3D {0};
> > > + Oid slot_database =3D local_slot->data.database;
> > >   bool synced_slot;
> > >
> > > Is it really safe to read slot_database before acquiring the database=
 lock?
> >
> > Reading slot_database before taking the database lock seems not
> > inherently unsafe by itself. The comment suggests that the lock is
> > primarily used to prevent conflicts with the startup process running
> > ReplicationSlotsDropDBSlots() during db-drop replay; it does not
> > protect replication slot array reuse.
> >
> > The unsafe part could be reading slot_database from local_slot after
> > ReplicationSlotControlLock has been released. At this point, the slot
> > array cell may already have been freed and reused, so the value read
> > may no longer belong to the slot that get_local_synced_slots()
> > originally collected. As a result, we could end up locking the wrong
> > database.
> >
> > There seems to be two related issues:
> >
> > 1) Before drop: reading local_slot->data.database /
> > local_slot->data.name after the slot-array lock was released, before
> > verifying the cell still represents the same synced slot.
>
> I recall condition (1) is considered acceptable, since the database lock =
is
> released immediately after re-verifying that the slot is no longer the or=
iginal
> 'synced' one anyway. Additionally, this race can only occur when replayin=
g a
> DROP DATABASE, which is rare in practice. Since we only take a shared loc=
k, it
> does not seem to cause real issues.
>

It seems that (1) is talking about the access to local_slot->data.name
before we acquire database lock in local_sync_slot_required() whereas
your response doesn't seem to address that concern. If not, then how
exactly does the database lock protect what we are doing in
local_sync_slot_required()?

--=20
With Regards,
Amit Kapila.