Re: [PATCH v2 net] dccp/tcp: Unhash sk from ehash for tb2 alloc failure after check_estalblished().

[Kuniyuki Iwashima <kuniyu@amazon.com>]

From: Eric Dumazet <edumazet@google.com>
Date: Wed, 14 Feb 2024 10:05:33 +0100
> On Tue, Feb 13, 2024 at 10:42 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> >
> > syzkaller reported a warning [0] in inet_csk_destroy_sock() with no
> > repro.
> >
> >   WARN_ON(inet_sk(sk)->inet_num && !inet_csk(sk)->icsk_bind_hash);
> >
> > However, the syzkaller's log hinted that connect() failed just before
> > the warning due to FAULT_INJECTION.  [1]
> >
> > When connect() is called for an unbound socket, we search for an
> > available ephemeral port.  If a bhash bucket exists for the port, we
> > call __inet_check_established() or __inet6_check_established() to check
> > if the bucket is reusable.
> >
> > If reusable, we add the socket into ehash and set inet_sk(sk)->inet_num.
> >
> > Later, we look up the corresponding bhash2 bucket and try to allocate
> > it if it does not exist.
> >
> > Although it rarely occurs in real use, if the allocation fails, we must
> > revert the changes by check_established().  Otherwise, an unconnected
> > socket could illegally occupy an ehash entry.
> >
> > Note that we do not put tw back into ehash because sk might have
> > already responded to a packet for tw and it would be better to free
> > tw earlier under such memory presure.
> >
> > [0]:
> >
> > Reported-by: syzkaller <syzkaller@googlegroups.com>
> > Fixes: 28044fc1d495 ("net: Add a bhash2 table hashed by port and address")
> > Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
> > ---
> > v2:
> >   * Unhash twsk from bhash/bhash2
> >
> > v1: https://lore.kernel.org/netdev/20240209025409.27235-1-kuniyu@amazon.com/
> > ---
> >  net/ipv4/inet_hashtables.c | 21 +++++++++++++++++++++
> >  1 file changed, 21 insertions(+)
> >
> > diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
> > index 93e9193df544..b22c71f93297 100644
> > --- a/net/ipv4/inet_hashtables.c
> > +++ b/net/ipv4/inet_hashtables.c
> > @@ -1130,10 +1130,31 @@ int __inet_hash_connect(struct inet_timewait_death_row *death_row,
> >         return 0;
> >
> >  error:
> > +       if (sk_hashed(sk)) {
> > +               spinlock_t *lock = inet_ehash_lockp(hinfo, sk->sk_hash);
> > +
> > +               sock_prot_inuse_add(net, sk->sk_prot, -1);
> > +
> > +               spin_lock(lock);
> > +               sk_nulls_del_node_init_rcu(sk);
> > +               spin_unlock(lock);
> > +
> > +               sk->sk_hash = 0;
> > +               inet_sk(sk)->inet_sport = 0;
> > +               inet_sk(sk)->inet_num = 0;
> > +
> > +               if (tw)
> > +                       inet_twsk_bind_unhash(tw, hinfo);
> > +       }
> > +
> >         spin_unlock(&head2->lock);
> >         if (tb_created)
> >                 inet_bind_bucket_destroy(hinfo->bind_bucket_cachep, tb);
> >         spin_unlock_bh(&head->lock);
> > +
> > +       if (tw)
> > +               inet_twsk_deschedule_put(tw);
> 
> Please make sure to call this while BH is still disabled.
> 
> 
>     spin_unlock(&head->lock);
>     if (tw)
>         inet_twsk_deschedule_put(tw);
>    local_bh_enable();

Sure, will update in v3 and post a followup for the existing
inet_twsk_deschedule_put() to net-next later.

Thanks!