Forwarded: [PATCH] Bluetooth: L2CAP: fix race l2cap_sock_cleanup_listen() vs. put_chan

[syzbot <syzbot+e6382a2f53f5fc7453ac@syzkaller.appspotmail.com>]

For archival purposes, forwarding an incoming command email to
linux-kernel@vger.kernel.org.

***

Subject: [PATCH] Bluetooth: L2CAP: fix race l2cap_sock_cleanup_listen() vs. put_chan
Author: pav@iki.fi

For L2CAP sockets without owning sk->sk_socket, reading
l2cap_pi(sk)->chan may race against concurrent l2cap_sock_kill() ->
l2cap_sock_put_chan().  This excludes simultaneous proto_ops callbacks,
but access in l2cap_sock_cleanup_listen() has unsafe lockless read.

 [Task 1]                         [Task 2 (hdev->workqueue)]
 l2cap_sock_release(parent)       l2cap_disconn_cfm
   l2cap_sock_cleanup_listen        l2cap_conn_del
     bt_accept_dequeue                l2cap_chan_del
       lock_sock(sk)                    l2cap_sock_teardown_cb
       bt_accept_unlink
         bt_sk(sk)->parent = NULL
       release_sock(sk) ----------------> lock_sock(sk)
                                          parent = /* NULL */
     lock_sock(sk) <--------------------- release_sock(sk)
                                          sock_set_flag(sk, SOCK_ZAPPED)
                                      l2cap_sock_close_cb
                                        l2cap_sock_kill(sk)
                                          l2cap_sock_put_chan
     chan = READ l2cap_pi(sk)->chan         l2cap_pi(sk)->chan = NULL
     l2cap_chan_hold_unless_zero            l2cap_put_chan(chan)
       kref_get_unless_zero(&chan->ref)

Task 1 may observe NULL which causes null-ptr-deref.

Fix the race by taking lock_sock() in l2cap_sock_kill() to
synchronize with l2cap_sock_cleanup_listen().  hold_unless_zero() is not
needed here, l2cap_pi(sk)->chan owns reference if it is non-NULL.

Clarify code comments vs. locking.

Fixes: 0e2c0392b9dc ("Bluetooth: L2CAP: Fix use-after-free in l2cap_sock_new_connection_cb()")
Reported-by: syzbot+e6382a2f53f5fc7453ac@syzkaller.appspotmail.com
---

#syz test

 include/net/bluetooth/l2cap.h |  5 +++++
 net/bluetooth/l2cap_sock.c    | 23 +++++++++++++----------
 2 files changed, 18 insertions(+), 10 deletions(-)

diff --git a/include/net/bluetooth/l2cap.h b/include/net/bluetooth/l2cap.h
index ef6ce1c20a4f..3d9a32094347 100644
--- a/include/net/bluetooth/l2cap.h
+++ b/include/net/bluetooth/l2cap.h
@@ -699,7 +699,12 @@ struct l2cap_rx_busy {
 
 struct l2cap_pinfo {
 	struct bt_sock		bt;
+
+	/* With owning sk_socket chan may be read without lock, other access
+	 * should hold lock_sock.
+	 */
 	struct l2cap_chan	*chan;
+
 	struct list_head	rx_busy;
 };
 
diff --git a/net/bluetooth/l2cap_sock.c b/net/bluetooth/l2cap_sock.c
index 735167f73f31..9540617a0e6c 100644
--- a/net/bluetooth/l2cap_sock.c
+++ b/net/bluetooth/l2cap_sock.c
@@ -1312,7 +1312,12 @@ static void l2cap_sock_kill(struct sock *sk)
 
 	BT_DBG("sk %p state %s", sk, state_to_string(sk->sk_state));
 
+	/* Take lock to synchronize against access without owning sk->sk_socket,
+	 * eg. in l2cap_sock_cleanup_listen(). proto_ops etc. don't need lock.
+	 */
+	lock_sock(sk);
 	l2cap_sock_put_chan(sk);
+	release_sock(sk);
 
 	/* Kill poor orphan */
 	sock_set_flag(sk, SOCK_DEAD);
@@ -1516,14 +1521,10 @@ static void l2cap_sock_cleanup_listen(struct sock *parent)
 	 * establish sk_lock -> conn->lock and invert the established
 	 * conn->lock -> chan->lock -> sk_lock order (lockdep deadlock).
 	 *
-	 * Instead, briefly take the child sk lock to fetch and pin its chan.
-	 * l2cap_conn_del() reaches the chan free only via
-	 * l2cap_chan_del() -> l2cap_sock_teardown_cb(), which itself takes
-	 * the child sk lock; holding it across l2cap_chan_hold_unless_zero()
-	 * therefore guarantees the chan cannot be freed while we read and
-	 * pin it (hold_unless_zero() additionally skips a chan already past
-	 * its last reference).  We then drop the sk lock before taking
-	 * chan->lock, so sk and chan locks are never held together.
+	 * Instead, briefly take the child sk lock to synchronize vs.
+	 * l2cap_sock_kill that puts l2cap_pi(sk)->chan. We then drop the sk
+	 * lock before taking chan->lock, so sk and chan locks are never held
+	 * together.
 	 *
 	 * Since we cannot call l2cap_chan_close() without conn->lock,
 	 * schedule l2cap_chan_timeout to close the channel; it already
@@ -1533,10 +1534,12 @@ static void l2cap_sock_cleanup_listen(struct sock *parent)
 		struct l2cap_chan *chan;
 
 		lock_sock_nested(sk, L2CAP_NESTING_NORMAL);
-		chan = l2cap_chan_hold_unless_zero(l2cap_pi(sk)->chan);
+		chan = l2cap_pi(sk)->chan;
+		if (chan)
+			l2cap_chan_hold(chan);
 		release_sock(sk);
 		if (!chan) {
-			/* l2cap_conn_del() already tearing this child down */
+			/* Already torn down */
 			sock_put(sk);
 			continue;
 		}
-- 
2.55.0