Syzbot reported an inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W} lock usage in sco_conn_del and sco_sock_timeout that could lead to deadlocks: https://syzkaller.appspot.com/bug?id=9089d89de0502e120f234ca0fc8a703f7368b31e This inconsistent lock state can also happen between sco_conn_ready and sco_sock_timeout. The issue is that these three functions take a spin lock on the socket, but sco_sock_timeout is called from an IRQ context. Since bh_lock_sock calls spin_lock but does not disable softirqs, this could lead to deadlocks: CPU0 ---- lock(slock-AF_BLUETOOTH-BTPROTO_SCO); <Interrupt> lock(slock-AF_BLUETOOTH-BTPROTO_SCO); *** DEADLOCK *** We fix this by replacing bh_lock_sock with spin_lock_bh in sco_conn_del and sco_conn_ready. Additionally, to avoid regressions, we pull the clean-up code out from sco_chan_del and use it directly in sco_conn_del. This is necessary because sco_chan_del makes a call to sco_conn_lock which takes an SOFTIRQ-unsafe lock. This means that calling sco_chan_del while holding the socket lock would result in a SOFTIRQ-safe -> SOFTIRQ-unsafe lock hierarchy between slock-AF_BLUETOOTH-BTPROTO_SCO and &conn->lock#2. This could lead to a deadlock as well: CPU0 CPU1 ---- ---- lock(&conn->lock#2); local_irq_disable(); lock(slock-AF_BLUETOOTH-BTPROTO_SCO); lock(&conn->lock#2); <Interrupt> lock(slock-AF_BLUETOOTH-BTPROTO_SCO); *** DEADLOCK *** Pulling out the code from sco_chan_del allows us to avoid this lock dependency by holding the two locks for only their required critical sections. Reported-and-tested-by: syzbot+2f6d7c28bb4bf7e82060@xxxxxxxxxxxxxxxxxxxxxxxxx Signed-off-by: Desmond Cheong Zhi Xi <desmondcheongzx@xxxxxxxxx> --- net/bluetooth/sco.c | 28 ++++++++++++++++++++-------- 1 file changed, 20 insertions(+), 8 deletions(-) diff --git a/net/bluetooth/sco.c b/net/bluetooth/sco.c index 3bd41563f118..d05629d7cc55 100644 --- a/net/bluetooth/sco.c +++ b/net/bluetooth/sco.c @@ -173,10 +173,22 @@ static void sco_conn_del(struct hci_conn *hcon, int err) if (sk) { sock_hold(sk); - bh_lock_sock(sk); + + spin_lock_bh(&sk->sk_lock.slock); sco_sock_clear_timer(sk); - sco_chan_del(sk, err); - bh_unlock_sock(sk); + sco_pi(sk)->conn = NULL; + if (conn->hcon) + hci_conn_drop(conn->hcon); + sk->sk_state = BT_CLOSED; + sk->sk_err = err; + sk->sk_state_change(sk); + sock_set_flag(sk, SOCK_ZAPPED); + spin_unlock_bh(&sk->sk_lock.slock); + + sco_conn_lock(conn); + conn->sk = NULL; + sco_conn_unlock(conn); + sco_sock_kill(sk); sock_put(sk); } @@ -1084,10 +1096,10 @@ static void sco_conn_ready(struct sco_conn *conn) if (sk) { sco_sock_clear_timer(sk); - bh_lock_sock(sk); + spin_lock_bh(&sk->sk_lock.slock); sk->sk_state = BT_CONNECTED; sk->sk_state_change(sk); - bh_unlock_sock(sk); + spin_unlock_bh(&sk->sk_lock.slock); } else { sco_conn_lock(conn); @@ -1102,12 +1114,12 @@ static void sco_conn_ready(struct sco_conn *conn) return; } - bh_lock_sock(parent); + spin_lock_bh(&parent->sk_lock.slock); sk = sco_sock_alloc(sock_net(parent), NULL, BTPROTO_SCO, GFP_ATOMIC, 0); if (!sk) { - bh_unlock_sock(parent); + spin_unlock_bh(&parent->sk_lock.slock); sco_conn_unlock(conn); return; } @@ -1128,7 +1140,7 @@ static void sco_conn_ready(struct sco_conn *conn) /* Wake up parent */ parent->sk_data_ready(parent); - bh_unlock_sock(parent); + spin_unlock_bh(&parent->sk_lock.slock); sco_conn_unlock(conn); } -- 2.25.1