Re: [PATCH net v3] mptcp: pm: Fix uaf in __timer_delete_sync

[Paolo Abeni <pabeni@redhat.com>]

On 9/5/24 14:27, Edward Adam Davis wrote:
> There are two paths to access mptcp_pm_del_add_timer, result in a race
> condition:
> 
>       CPU1				CPU2
>       ====                               ====
>       net_rx_action
>       napi_poll                          netlink_sendmsg
>       __napi_poll                        netlink_unicast
>       process_backlog                    netlink_unicast_kernel
>       __netif_receive_skb                genl_rcv
>       __netif_receive_skb_one_core       netlink_rcv_skb
>       NF_HOOK                            genl_rcv_msg
>       ip_local_deliver_finish            genl_family_rcv_msg
>       ip_protocol_deliver_rcu            genl_family_rcv_msg_doit
>       tcp_v4_rcv                         mptcp_pm_nl_flush_addrs_doit
>       tcp_v4_do_rcv                      mptcp_nl_remove_addrs_list
>       tcp_rcv_established                mptcp_pm_remove_addrs_and_subflows
>       tcp_data_queue                     remove_anno_list_by_saddr
>       mptcp_incoming_options             mptcp_pm_del_add_timer
>       mptcp_pm_del_add_timer             kfree(entry)
> 
> In remove_anno_list_by_saddr(running on CPU2), after leaving the critical
> zone protected by "pm.lock", the entry will be released, which leads to the
> occurrence of uaf in the mptcp_pm_del_add_timer(running on CPU1).
> 
> Keeping a reference to add_timer inside the lock, and calling
> sk_stop_timer_sync() with this reference, instead of "entry->add_timer".
> 
> Fixes: 00cfd77b9063 ("mptcp: retransmit ADD_ADDR when timeout")
> Cc: stable@vger.kernel.org
> Reported-and-tested-by: syzbot+f3a31fb909db9b2a5c4d@syzkaller.appspotmail.com
> Closes: https://syzkaller.appspot.com/bug?extid=f3a31fb909db9b2a5c4d
> Signed-off-by: Matthieu Baerts (NGI0) <matttbe@kernel.org>
> Signed-off-by: Edward Adam Davis <eadavis@qq.com>
> ---
>   net/mptcp/pm_netlink.c | 14 ++++++++++----
>   1 file changed, 10 insertions(+), 4 deletions(-)
> 
> diff --git a/net/mptcp/pm_netlink.c b/net/mptcp/pm_netlink.c
> index 3e4ad801786f..7ddb373cc6ad 100644
> --- a/net/mptcp/pm_netlink.c
> +++ b/net/mptcp/pm_netlink.c
> @@ -329,17 +329,21 @@ struct mptcp_pm_add_entry *
>   mptcp_pm_del_add_timer(struct mptcp_sock *msk,
>   		       const struct mptcp_addr_info *addr, bool check_id)
>   {
> -	struct mptcp_pm_add_entry *entry;
>   	struct sock *sk = (struct sock *)msk;
> +	struct timer_list *add_timer = NULL;
> +	struct mptcp_pm_add_entry *entry;
>   
>   	spin_lock_bh(&msk->pm.lock);
>   	entry = mptcp_lookup_anno_list_by_saddr(msk, addr);
> -	if (entry && (!check_id || entry->addr.id == addr->id))
> +	if (entry && (!check_id || entry->addr.id == addr->id)) {
>   		entry->retrans_times = ADD_ADDR_RETRANS_MAX;
> +		add_timer = &entry->add_timer;
> +	}
>   	spin_unlock_bh(&msk->pm.lock);
>   
> -	if (entry && (!check_id || entry->addr.id == addr->id))
> -		sk_stop_timer_sync(sk, &entry->add_timer);
> +	/* no lock, because sk_stop_timer_sync() is calling del_timer_sync() */
> +	if (add_timer)
> +		sk_stop_timer_sync(sk, add_timer);
>   
>   	return entry;
>   }
> @@ -1430,8 +1434,10 @@ static bool remove_anno_list_by_saddr(struct mptcp_sock *msk,
>   
>   	entry = mptcp_pm_del_add_timer(msk, addr, false);
>   	if (entry) {
> +		spin_lock_bh(&msk->pm.lock);
>   		list_del(&entry->list);
>   		kfree(entry);
> +		spin_unlock_bh(&msk->pm.lock);

I'm sorry for the late feedback.

I think this is not enough to fix races for good, i.e.

mptcp_nl_remove_subflow_and_signal_addr() -> mptcp_pm_remove_anno_addr()
-> remove_anno_list_by_saddr()

could race with:

mptcp_pm_remove_addrs() -> remove_anno_list_by_saddr()

and both CPUs could see the same 'entry' returned by
mptcp_pm_del_add_timer().

I think the list_del() in remove_anno_list_by_saddr() should moved under
the pm lock protection inside mptcp_pm_del_add_timer(), and no need to 
add spin_lock_bh() around the kfree call.

Thanks,

Paolo