Bluetooth: avoid hci_dev_test_and_set_flag() in mgmt_init_hdev()

syzbot is again reporting attempt to cancel uninitialized work
at mgmt_index_removed() [1], for setting of HCI_MGMT flag from
mgmt_init_hdev() from hci_mgmt_cmd() from hci_sock_sendmsg() can
race with testing of HCI_MGMT flag from mgmt_index_removed() from
hci_sock_bind() due to lack of serialization via hci_dev_lock().

Since mgmt_init_hdev() is called with mgmt_chan_list_lock held, we can
safely split hci_dev_test_and_set_flag() into hci_dev_test_flag() and
hci_dev_set_flag(). Thus, in order to close this race, set HCI_MGMT flag
after INIT_DELAYED_WORK() completed.

This is a local fix based on mgmt_chan_list_lock. Lack of serialization
via hci_dev_lock() might be causing different race conditions somewhere
else. But a global fix based on hci_dev_lock() should deserve a future
patch.

Link: https://syzkaller.appspot.com/bug?extid=844c7bf1b1aa4119c5de
Reported-by: syzbot+844c7bf1b1aa4119c5de@syzkaller.appspotmail.com
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Fixes: 3f2893d3c1 ("Bluetooth: don't try to cancel uninitialized works at mgmt_index_removed()")
Signed-off-by: Luiz Augusto von Dentz <luiz.von.dentz@intel.com>

net/bluetooth/mgmt.c

@@ -1111,7 +1111,7 @@ static void mesh_send_done(struct work_struct *work)
 
 static void mgmt_init_hdev(struct sock *sk, struct hci_dev *hdev)
 {
-	if (hci_dev_test_and_set_flag(hdev, HCI_MGMT))
+	if (hci_dev_test_flag(hdev, HCI_MGMT))
 		return;
 
 	BT_INFO("MGMT ver %d.%d", MGMT_VERSION, MGMT_REVISION);
@@ -1127,6 +1127,8 @@ static void mgmt_init_hdev(struct sock *sk, struct hci_dev *hdev)
 	 * it
 	 */
 	hci_dev_clear_flag(hdev, HCI_BONDABLE);
+
+	hci_dev_set_flag(hdev, HCI_MGMT);
 }
 
 static int read_controller_info(struct sock *sk, struct hci_dev *hdev,