Re: [PATCH bpf-next 1/2] sk_msg: Keep reference on socket file while wait_memory

[Jakub Sitnicki <jakub@cloudflare.com>]

On Fri, Aug 19, 2022 at 10:01 AM GMT, liujian (CE) wrote:
>> -----Original Message-----
>> From: Jakub Sitnicki [mailto:jakub@cloudflare.com]
>> Sent: Friday, August 19, 2022 4:39 PM
>> To: liujian (CE) <liujian56@huawei.com>; john.fastabend@gmail.com;
>> edumazet@google.com
>> Cc: davem@davemloft.net; yoshfuji@linux-ipv6.org; dsahern@kernel.org;
>> kuba@kernel.org; pabeni@redhat.com; andrii@kernel.org; mykolal@fb.com;
>> ast@kernel.org; daniel@iogearbox.net; martin.lau@linux.dev;
>> song@kernel.org; yhs@fb.com; kpsingh@kernel.org; sdf@google.com;
>> haoluo@google.com; jolsa@kernel.org; shuah@kernel.org;
>> bpf@vger.kernel.org
>> Subject: Re: [PATCH bpf-next 1/2] sk_msg: Keep reference on socket file
>> while wait_memory
>> 
>> On Mon, Aug 15, 2022 at 10:33 AM +08, Liu Jian wrote:
>> > Fix the below NULL pointer dereference:
>> >
>> > [   14.471200] Call Trace:
>> > [   14.471562]  <TASK>
>> > [   14.471882]  lock_acquire+0x245/0x2e0
>> > [   14.472416]  ? remove_wait_queue+0x12/0x50
>> > [   14.473014]  ? _raw_spin_lock_irqsave+0x17/0x50
>> > [   14.473681]  _raw_spin_lock_irqsave+0x3d/0x50
>> > [   14.474318]  ? remove_wait_queue+0x12/0x50
>> > [   14.474907]  remove_wait_queue+0x12/0x50
>> > [   14.475480]  sk_stream_wait_memory+0x20d/0x340
>> > [   14.476127]  ? do_wait_intr_irq+0x80/0x80
>> > [   14.476704]  do_tcp_sendpages+0x287/0x600
>> > [   14.477283]  tcp_bpf_push+0xab/0x260
>> > [   14.477817]  tcp_bpf_sendmsg_redir+0x297/0x500
>> > [   14.478461]  ? __local_bh_enable_ip+0x77/0xe0
>> > [   14.479096]  tcp_bpf_send_verdict+0x105/0x470
>> > [   14.479729]  tcp_bpf_sendmsg+0x318/0x4f0
>> > [   14.480311]  sock_sendmsg+0x2d/0x40
>> > [   14.480822]  ____sys_sendmsg+0x1b4/0x1c0
>> > [   14.481390]  ? copy_msghdr_from_user+0x62/0x80
>> > [   14.482048]  ___sys_sendmsg+0x78/0xb0
>> > [   14.482580]  ? vmf_insert_pfn_prot+0x91/0x150
>> > [   14.483215]  ? __do_fault+0x2a/0x1a0
>> > [   14.483738]  ? do_fault+0x15e/0x5d0
>> > [   14.484246]  ? __handle_mm_fault+0x56b/0x1040
>> > [   14.484874]  ? lock_is_held_type+0xdf/0x130
>> > [   14.485474]  ? find_held_lock+0x2d/0x90
>> > [   14.486046]  ? __sys_sendmsg+0x41/0x70
>> > [   14.486587]  __sys_sendmsg+0x41/0x70
>> > [   14.487105]  ? intel_pmu_drain_pebs_core+0x350/0x350
>> > [   14.487822]  do_syscall_64+0x34/0x80
>> > [   14.488345]  entry_SYSCALL_64_after_hwframe+0x63/0xcd
>> >
>> > The test scene as following flow:
>> > thread1                               thread2
>> > -----------                           ---------------
>> >  tcp_bpf_sendmsg
>> >   tcp_bpf_send_verdict
>> >    tcp_bpf_sendmsg_redir              sock_close
>> >     tcp_bpf_push_locked                 __sock_release
>> >      tcp_bpf_push                         //inet_release
>> >       do_tcp_sendpages                    sock->ops->release
>> >        sk_stream_wait_memory          	   // tcp_close
>> >           sk_wait_event                      sk->sk_prot->close
>> >            release_sock(__sk);
>> >             ***
>> >
>> >                                                 lock_sock(sk);
>> >                                                   __tcp_close
>> >                                                     sock_orphan(sk)
>> >                                                       sk->sk_wq  = NULL
>> >                                                 release_sock
>> >             ****
>> >            lock_sock(__sk);
>> >           remove_wait_queue(sk_sleep(sk), &wait);
>> >              sk_sleep(sk)
>> >              //NULL pointer dereference
>> >              &rcu_dereference_raw(sk->sk_wq)->wait
>> >
>> > While waiting for memory in thread1, the socket is released with its
>> > wait queue because thread2 has closed it. This caused by
>> > tcp_bpf_send_verdict didn't increase the f_count of psock->sk_redir-
>> >sk_socket->file in thread1.
>> 
>> I'm not sure about this approach. Keeping a closed sock file alive, just so we
>> can wakeup from sleep, seems like wasted effort.
>> 
>> __tcp_close sets sk->sk_shutdown = RCV_SHUTDOWN | SEND_SHUTDOWN.
>> So we will return from sk_stream_wait_memory via the do_error path.
>> 
>> SEND_SHUTDOWN might be set because socket got closed and orphaned -
>> dead and detached from its file, like in this case.
>> 
>> So, IMHO, we should check if SOCK_DEAD flag is set on wakeup due to
>> SEND_SHUTDOWN in sk_stream_wait_memory, before accessing the wait
>> queue.
>> 
>> [...]
> As jakub's approach, this problem can be solved.
>
> diff --git a/include/net/sock.h b/include/net/sock.h
> index a7273b289188..a3dab7140f1e 100644
> --- a/include/net/sock.h
> +++ b/include/net/sock.h
> @@ -1998,6 +1998,8 @@ static inline void sk_set_socket(struct sock *sk, struct socket *sock)
>  static inline wait_queue_head_t *sk_sleep(struct sock *sk)
>  {
>         BUILD_BUG_ON(offsetof(struct socket_wq, wait) != 0);
> +       if (sock_flag(sk, SOCK_DEAD))
> +               return NULL;
>         return &rcu_dereference_raw(sk->sk_wq)->wait;
>  }
>  /* Detach socket from process context.
> diff --git a/kernel/sched/wait.c b/kernel/sched/wait.c
> index 9860bb9a847c..da1be17d0b19 100644
> --- a/kernel/sched/wait.c
> +++ b/kernel/sched/wait.c
> @@ -51,6 +51,8 @@ void remove_wait_queue(struct wait_queue_head *wq_head, struct wait_queue_entry
>  {
>         unsigned long flags;
>  
> +       if (wq_head == NULL)
> +               return;
>         spin_lock_irqsave(&wq_head->lock, flags);
>         __remove_wait_queue(wq_head, wq_entry);
>         spin_unlock_irqrestore(&wq_head->lock, flags);

I don't know if we want to change the contract for sk_sleep()
remove_wait_queue() so that they accept dead sockets or nulls.

How about just:

diff --git a/net/core/stream.c b/net/core/stream.c
index ccc083cdef23..1105057ce00a 100644
--- a/net/core/stream.c
+++ b/net/core/stream.c
@@ -159,7 +159,8 @@ int sk_stream_wait_memory(struct sock *sk, long *timeo_p)
                *timeo_p = current_timeo;
        }
 out:
-       remove_wait_queue(sk_sleep(sk), &wait);
+       if (!sock_flag(sk, SOCK_DEAD))
+               remove_wait_queue(sk_sleep(sk), &wait);
        return err;

 do_error: