Re: [PATCH net] l2tp: use rcu_dereference_sk_user_data() in l2tp_udp_encap_recv()

[Guillaume Nault <gnault@redhat.com>]

On Wed, Apr 24, 2019 at 04:33:47AM -0700, Eric Dumazet wrote:
> On Wed, Apr 24, 2019 at 2:58 AM Guillaume Nault <gnault@redhat.com> wrote:
> >
> > On Tue, Apr 23, 2019 at 09:43:26AM -0700, Eric Dumazet wrote:
> > > Canonical way to fetch sk_user_data from an encap_rcv() handler called
> > > from UDP stack in rcu protected section is to use rcu_dereference_sk_user_data(),
> > > otherwise compiler might read it multiple times.
> > >
> > That reminds me the more general problem we have with ->sk_user_data:
> > https://lore.kernel.org/netdev/20180117.142538.1972806008716856078.davem@davemloft.net/
> >
> > We're not even guarateed that ->sk_user_data points to a struct l2tp_tunnel
> > (some external modules can still probably override it).
> >
> 
> RCU rules should prevail.
> 
> An RCU grace period must be enforced if the same UDP socket has its
> sk_user_data changed.
> 
> Normally, UDP socket observes an RCU grace period at close/destroy
> time. (SOCK_RCU_FREE)
> 
> If we detect the same UDP socket has not been closed between clearing
> sk_user_data
> and setting it again to a non NULL value, we must insert a synchronize_rcu()
> 
> A generic change could look like the following (but we must check that all
> rcu_assign_sk_user_data() callers can sleep)
> 
It seems that callers of rcu_assign_sk_user_data() can indeed sleep.

But I think we have more problems. Not all code paths treat
->sk_user_data as RCU pointers (IIUC that's why we created the
__sk_user_data() macro, instead of just redefining ->sk_user_data as
"void __rcu *"). But even if RCU rules were respected for all accesses,
we'd need to ensure consistent protection for the update side.

And then, we'd need to make sure that ->sk_user_data is in sync with
the encap_rcv() callback (or whatever actually uses the data pointed
to). Otherwise a module could treat ->sk_user_data as a struct foo
pointer while it actually points to a struct bar.

For example, a quick look at net/sunrpc/svcsock.c seems to indicate
that svc_addsock() would accept any (unconnected) UDP socket and pass
it to svc_addsock(), which in turn would override ->sk_user_data with
a struct svc_sock pointer. If the socket was previously set up by L2TP,
then we'd end up with ->sk_user_data pointing to a svc_sock structure,
but ->encap_rcv still pointing to l2tp_udp_encap_recv(). That's going
to give unexpected results when l2tp_udp_encap_recv() will dereference
->sk_user_data to access (what it believes to be) its tunnel structure.

> diff --git a/include/net/sock.h b/include/net/sock.h
> index 784cd19d5ff76b53865f79d4580f3fa269fa2408..3cf0dfd9d83a956220d69138339561b62407addd
> 100644
> --- a/include/net/sock.h
> +++ b/include/net/sock.h
> @@ -522,7 +522,6 @@ enum sk_pacing {
>  #define __sk_user_data(sk) ((*((void __rcu **)&(sk)->sk_user_data)))
> 
>  #define rcu_dereference_sk_user_data(sk)
> rcu_dereference(__sk_user_data((sk)))
> -#define rcu_assign_sk_user_data(sk, ptr)
> rcu_assign_pointer(__sk_user_data((sk)), ptr)
> 
>  /*
>   * SK_CAN_REUSE and SK_NO_REUSE on a socket mean that the socket is OK
> @@ -811,6 +810,7 @@ enum sock_flags {
>         SOCK_FILTER_LOCKED, /* Filter cannot be changed anymore */
>         SOCK_SELECT_ERR_QUEUE, /* Wake select on error queue */
>         SOCK_RCU_FREE, /* wait rcu grace period in sk_destruct() */
> +       SOCK_USER_DATA_SET,
>         SOCK_TXTIME,
>         SOCK_XDP, /* XDP is attached */
>         SOCK_TSTAMP_NEW, /* Indicates 64 bit timestamps always */
> @@ -2582,4 +2582,15 @@ static inline bool sk_dev_equal_l3scope(struct
> sock *sk, int dif)
>         return false;
>  }
> 
> +static inline void rcu_assign_sk_user_data(struct sock *sk, void *ptr)
> +{
> +       if (ptr) {
> +               might_sleep();
> +               if (sock_flag(sk, SOCK_USER_DATA_SET))
> +                       synchronize_net();
> +       }
> +       rcu_assign_pointer(__sk_user_data((sk)), ptr);
> +       if (ptr)
> +               sock_set_flag(sk, SOCK_USER_DATA_SET);
> +}
>  #endif /* _SOCK_H */
> 
> 
> 
> > There were some locking rules defined for setting ->sk_user_data:
> > https://lore.kernel.org/netdev/20180124203541.3172-3-tom@quantonium.net/
> > Converting all users to either avoid using ->sk_user_data or using it
> > under the proper pre-conditions has been on my TODO list for a while...
> > I'll try to revive this effort.
> 
> Avoiding using sk_user_data seems not practical.
> 
> However making sure it is not blindly overwritten by a buggy module is doable.
> 
At least for those that use rcu_assign_sk_user_data(). And even then,
we'd need to ensure callers do properly validate the socket, to avoid
overriding ->sk_user_data with a different pointer type than what other
users expect.

> >
> > > Fixes: d00fa9adc528 ("il2tp: fix races with tunnel socket close")
> >                         ^
> > Spurious "i". Vi user? :)
> 
> I am not a vi user ;)
> 
So nice to hear :)

> >
> > > Signed-off-by: Eric Dumazet <edumazet@google.com>
> > > Cc: James Chapman <jchapman@katalix.com>
> > > ---
> > >  net/l2tp/l2tp_core.c | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > >
> > > diff --git a/net/l2tp/l2tp_core.c b/net/l2tp/l2tp_core.c
> > > index fed6becc5daf86afa2ad9188bb28e151244bb5a6..aee33d1320184e411dbedff72b5bf5199481e53f 100644
> > > --- a/net/l2tp/l2tp_core.c
> > > +++ b/net/l2tp/l2tp_core.c
> > > @@ -909,7 +909,7 @@ int l2tp_udp_encap_recv(struct sock *sk, struct sk_buff *skb)
> > >  {
> > >       struct l2tp_tunnel *tunnel;
> > >
> > > -     tunnel = l2tp_tunnel(sk);
> > > +     tunnel = rcu_dereference_sk_user_data(sk);
> > >       if (tunnel == NULL)
> > >               goto pass_up;
> > >
> > > --
> > > 2.21.0.593.g511ec345e18-goog
> > >