Re: [PATCH v1 net] udp: Set SOCK_RCU_FREE earlier in udp_lib_get_port().

[Kuniyuki Iwashima <kuniyu@amazon.com>]

From: Eric Dumazet <edumazet@google.com>
Date: Mon, 8 Jul 2024 12:07:56 -0700
> On Mon, Jul 8, 2024 at 11:55 AM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> >
> > From: Eric Dumazet <edumazet@google.com>
> > Date: Mon, 8 Jul 2024 11:38:41 -0700
> > > On Mon, Jul 8, 2024 at 11:20 AM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> > > >
> > > > syzkaller triggered the warning [0] in udp_v4_early_demux().
> > > >
> > > > In udp_v4_early_demux(), we do not touch the refcount of the looked-up
> > > > sk and use sock_pfree() as skb->destructor, so we check SOCK_RCU_FREE
> > > > to ensure that the sk is safe to access during the RCU grace period.
> > > >
> > > > Currently, SOCK_RCU_FREE is flagged for a bound socket after being put
> > > > into the hash table.  Moreover, the SOCK_RCU_FREE check is done too
> > > > early in udp_v4_early_demux(), so there could be a small race window:
> > > >
> > > >   CPU1                                 CPU2
> > > >   ----                                 ----
> > > >   udp_v4_early_demux()                 udp_lib_get_port()
> > > >   |                                    |- hlist_add_head_rcu()
> > > >   |- sk = __udp4_lib_demux_lookup()    |
> > > >   |- DEBUG_NET_WARN_ON_ONCE(sk_is_refcounted(sk));
> > > >                                        `- sock_set_flag(sk, SOCK_RCU_FREE)
> > > >
> > > > In practice, sock_pfree() is called much later, when SOCK_RCU_FREE
> > > > is most likely propagated for other CPUs; otherwise, we will see
> > > > another warning of sk refcount underflow, but at least I didn't.
> > > >
> > > > Technically, moving sock_set_flag(sk, SOCK_RCU_FREE) before
> > > > hlist_add_{head,tail}_rcu() does not guarantee the order, and we
> > > > must put smp_mb() between them, or smp_wmb() there and smp_rmb()
> > > > in udp_v4_early_demux().
> > > >
> > > > But it's overkill in the real scenario, so I just put smp_mb() only under
> > > > CONFIG_DEBUG_NET to silence the splat.  When we see the refcount underflow
> > > > warning, we can remove the config guard.
> > > >
> > > > Another option would be to remove DEBUG_NET_WARN_ON_ONCE(), but this could
> > > > make future debugging harder without the hints in udp_v4_early_demux() and
> > > > udp_lib_get_port().
> > > >
> > > > [0]:
> > > >
> > > > Fixes: 08842c43d016 ("udp: no longer touch sk->sk_refcnt in early demux")
> > > > Reported-by: syzkaller <syzkaller@googlegroups.com>
> > > > Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
> > > > ---
> > > >  net/ipv4/udp.c | 8 +++++++-
> > > >  1 file changed, 7 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
> > > > index 189c9113fe9a..1a05cc3d2b4f 100644
> > > > --- a/net/ipv4/udp.c
> > > > +++ b/net/ipv4/udp.c
> > > > @@ -326,6 +326,12 @@ int udp_lib_get_port(struct sock *sk, unsigned short snum,
> > > >                         goto fail_unlock;
> > > >                 }
> > > >
> > > > +               sock_set_flag(sk, SOCK_RCU_FREE);
> > >
> > > Nice catch.
> > >
> > > > +
> > > > +               if (IS_ENABLED(CONFIG_DEBUG_NET))
> > > > +                       /* for DEBUG_NET_WARN_ON_ONCE() in udp_v4_early_demux(). */
> > > > +                       smp_mb();
> > > > +
> > >
> > > I do not think this smp_mb() is needed. If this was, many other RCU
> > > operations would need it,
> > >
> > > RCU rules mandate that all memory writes must be committed before the
> > > object can be seen
> > > by other cpus in the hash table.
> > >
> > > This includes the setting of the SOCK_RCU_FREE flag.
> > >
> > > For instance, hlist_add_head_rcu() does a
> > > rcu_assign_pointer(hlist_first_rcu(h), n);
> >
> > Ah, I was thinking spinlock will not prevent reordering, but
> > now I see, rcu_assign_pointer() had necessary barrier. :)
> >
> >   /**
> >    * rcu_assign_pointer() - assign to RCU-protected pointer
> >    ...
> >    * Assigns the specified value to the specified RCU-protected
> >    * pointer, ensuring that any concurrent RCU readers will see
> >    * any prior initialization.
> >
> > will remove smp_mb() and update the changelog in v2.
> >
> 
> A similar commit was
> 
> commit 871019b22d1bcc9fab2d1feba1b9a564acbb6e99
> Author: Stanislav Fomichev <sdf@fomichev.me>
> Date:   Wed Nov 8 13:13:25 2023 -0800
> 
>     net: set SOCK_RCU_FREE before inserting socket into hashtable
> 
> So I wonder if the bug could be older...

If we focus on the ordering, the Fixes tag would be

Fixes: ca065d0cf80f ("udp: no longer use SLAB_DESTROY_BY_RCU")

But, at that time, we had atomic_inc_not_zero_hint() and used
sock_efree(), which were removed later in 08842c43d016.

Which one should I use as Fixes: ?