bluetooth: fix locking bug in the rfcomm socket cleanup handling

in net/bluetooth/rfcomm/sock.c, rfcomm_sk_state_change() does the
following operation:

        if (parent && sock_flag(sk, SOCK_ZAPPED)) {
                /* We have to drop DLC lock here, otherwise
                 * rfcomm_sock_destruct() will dead lock. */
                rfcomm_dlc_unlock(d);
                rfcomm_sock_kill(sk);
                rfcomm_dlc_lock(d);
        }
}

which is fine, since rfcomm_sock_kill() will call sk_free() which will call
rfcomm_sock_destruct() which takes the rfcomm_dlc_lock()... so far so good.

HOWEVER, this assumes that the rfcomm_sk_state_change() function always gets
called with the rfcomm_dlc_lock() taken. This is the case for all but one
case, and in that case where we don't have the lock, we do a double unlock
followed by an attempt to take the lock, which due to underflow isn't
going anywhere fast.

This patch fixes this by moving the stragling case inside the lock, like
the other usages of the same call are doing in this code.

This was found with the help of the www.kerneloops.org project, where this
deadlock was observed 51 times at this point in time:
http://www.kerneloops.org/search.php?search=rfcomm_sock_destruct

Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Acked-by: Marcel Holtmann <marcel@holtmann.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

diff --git a/net/bluetooth/rfcomm/core.c b/net/bluetooth/rfcomm/core.c
index eb62558e9b092593269878c23379906cd6985873..0c2c93735e935417314e30eddbadc017222612ba 100644 (file)
--- a/net/bluetooth/rfcomm/core.c
+++ b/net/bluetooth/rfcomm/core.c
@@ -423,8 +423,8 @@ static int __rfcomm_dlc_close(struct rfcomm_dlc *d, int err)
 
                rfcomm_dlc_lock(d);
                d->state = BT_CLOSED;
-               rfcomm_dlc_unlock(d);
                d->state_change(d, err);
+               rfcomm_dlc_unlock(d);
 
                skb_queue_purge(&d->tx_queue);
                rfcomm_dlc_unlink(d);