Hello Gustavo,

The problem appears in case of multiple connect-transfer-disconnect sequence (e.g. by using l2test). The conditions are the following:
There are 2 BT devices. The first one listens and receives (l2test -r), the second one makes "connect-disconnect-connect..." sequence (l2test -c -b 1000 -i hci0 -P 10 <addr>). After some time this will cause the race between functions bt_accept_dequeue and l2cap_chan_del. The fail sequence:

struct sock *bt_accept_dequeue(struct sock *parent, struct socket *newsock)
{
...
    list_for_each_safe(p, n, &bt_sk(parent)->accept_q) {
       sk = (struct sock *) list_entry(p, struct bt_sock, accept_q);

       lock_sock(sk);

In this time the function l2cap_chan_del sets the socket state to BT_CLOSED, unlinks and kills the socket.

       /* FIXME: Is this check still needed */
       if (sk->sk_state == BT_CLOSED) {
           release_sock(sk);
           bt_accept_unlink(sk);
           continue;
       }

...

        release_sock(sk);
    }
    return NULL;
}

Regards,
Yuri

ext Gustavo F. Padovan wrote:

Hi Yuri,

* Yuri Ershov <ext-yuri.ershov@nokia.com> [2010-10-21 20:08:58 +0400]:

This patch fixes NULL pointer dereference at running test with
connect-transfer-disconnect in loop. Sometimes sk_state is 
BT_CLOSED and sk_refcnt equal to 0, so there is oops in 
bt_accept_unlink. In normal case removed block is not used.


Question here is: Why sk_refcnt is 0 at that point of the code?  The
socket should be destroyed if it ref is 0, but it wasn't, so something
in another point of the code went is wrong. "Sometimes" is not a good
description of the problem, you have to show why that happened.