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 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.

Thanks !