Re: [PATCH v3] net: ax25: Fix deadlock caused by skb_recv_datagram in ax25_recvmsg

[Paolo Abeni]

On Thu, 2022-06-09 at 21:17 +0800, duoming@xxxxxxxxxx wrote:
> Hello,
> 
> On Thu, 09 Jun 2022 10:41:02 +0200 Paolo wrote:
> 
> > On Wed, 2022-06-08 at 09:29 +0800, Duoming Zhou wrote:
> > > The skb_recv_datagram() in ax25_recvmsg() will hold lock_sock
> > > and block until it receives a packet from the remote. If the client
> > > doesn`t connect to server and calls read() directly, it will not
> > > receive any packets forever. As a result, the deadlock will happen.
> > > 
> > > The fail log caused by deadlock is shown below:
> > > 
> > > [  369.606973] INFO: task ax25_deadlock:157 blocked for more than 245 seconds.
> > > [  369.608919] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> > > [  369.613058] Call Trace:
> > > [  369.613315]  <TASK>
> > > [  369.614072]  __schedule+0x2f9/0xb20
> > > [  369.615029]  schedule+0x49/0xb0
> > > [  369.615734]  __lock_sock+0x92/0x100
> > > [  369.616763]  ? destroy_sched_domains_rcu+0x20/0x20
> > > [  369.617941]  lock_sock_nested+0x6e/0x70
> > > [  369.618809]  ax25_bind+0xaa/0x210
> > > [  369.619736]  __sys_bind+0xca/0xf0
> > > [  369.620039]  ? do_futex+0xae/0x1b0
> > > [  369.620387]  ? __x64_sys_futex+0x7c/0x1c0
> > > [  369.620601]  ? fpregs_assert_state_consistent+0x19/0x40
> > > [  369.620613]  __x64_sys_bind+0x11/0x20
> > > [  369.621791]  do_syscall_64+0x3b/0x90
> > > [  369.622423]  entry_SYSCALL_64_after_hwframe+0x46/0xb0
> > > [  369.623319] RIP: 0033:0x7f43c8aa8af7
> > > [  369.624301] RSP: 002b:00007f43c8197ef8 EFLAGS: 00000246 ORIG_RAX: 0000000000000031
> > > [  369.625756] RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007f43c8aa8af7
> > > [  369.626724] RDX: 0000000000000010 RSI: 000055768e2021d0 RDI: 0000000000000005
> > > [  369.628569] RBP: 00007f43c8197f00 R08: 0000000000000011 R09: 00007f43c8198700
> > > [  369.630208] R10: 0000000000000000 R11: 0000000000000246 R12: 00007fff845e6afe
> > > [  369.632240] R13: 00007fff845e6aff R14: 00007f43c8197fc0 R15: 00007f43c8198700
> > > 
> > > This patch moves the skb_recv_datagram() before lock_sock() in order that
> > > other functions that need lock_sock could be executed. What`s more, we
> > > add skb_free_datagram() before goto out in order to mitigate memory leak.
> > > 
> > > Suggested-by: Thomas Osterried <thomas@xxxxxxxxxxxx>
> > > Signed-off-by: Duoming Zhou <duoming@xxxxxxxxxx>
> > > Reported-by: Thomas Habets <thomas@@habets.se>
> > > ---
> > > Changes in v3:
> > >   - Add skb_free_datagram() before goto out in order to mitigate memory leak.
> > > 
> > >  net/ax25/af_ax25.c | 12 +++++++-----
> > >  1 file changed, 7 insertions(+), 5 deletions(-)
> > > 
> > > diff --git a/net/ax25/af_ax25.c b/net/ax25/af_ax25.c
> > > index 95393bb2760..62aa5993093 100644
> > > --- a/net/ax25/af_ax25.c
> > > +++ b/net/ax25/af_ax25.c
> > > @@ -1665,6 +1665,11 @@ static int ax25_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
> > >  	int copied;
> > >  	int err = 0;
> > >  
> > > +	/* Now we can treat all alike */
> > > +	skb = skb_recv_datagram(sk, flags, &err);
> > > +	if (!skb)
> > > +		goto done;
> > > +
> > 
> > Note that this causes a behavior change: before this patch, calling
> > recvmsg() on unconnected seqpacket sockets returned immediatelly with
> > an error (due to the the check below), now it blocks. 
> > 
> > The change may confuse (== break) user-space applications. I think it
> > would be better replacing skb_recv_datagram with an open-coded variant
> > of it releasing the socket lock before the
> > __skb_wait_for_more_packets() call and re-acquiring it after such call.
> > Somewhat alike __unix_dgram_recvmsg().
> 
> Thank you for your time and suggestions!
> I think the following method may solve the problem.
> 
> diff --git a/net/ax25/af_ax25.c b/net/ax25/af_ax25.c
> index 95393bb2760..51b441c837c 100644
> --- a/net/ax25/af_ax25.c
> +++ b/net/ax25/af_ax25.c
> @@ -1675,8 +1675,10 @@ static int ax25_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
>                 goto out;
>         }
> 
> +       release_sock(sk);
>         /* Now we can treat all alike */
>         skb = skb_recv_datagram(sk, flags, &err);
> +       lock_sock(sk);
>         if (skb == NULL)
>                 goto out;
> 
> The skb_recv_datagram() is free of race conditions and could be re-entrant.
> So calling skb_recv_datagram() without the protection of lock_sock() is ok.
> 
> What's more, releasing the lock_sock() before skb_recv_datagram() will not
> cause UAF bugs. Because the sock will not be deallocated unless we call
> ax25_release(), but ax25_release() and ax25_recvmsg() could not run in parallel.
> 
> Although the "sk->sk_state" may be changed due to the release of lock_sock(),
> it will not influence the following operations in ax25_recvmsg().

One of the downside of the above is that recvmsg() will unconditionally
acquire and release the socket lock twice which can have non
trivial/nasty side effects on process scheduling.

With the suggested change the socket lock will be released only when
recvmsg will block and that should produce nicer overal behavior.

Cheers,

Paolo